├── Configs
    └── config.yml
├── Data
    ├── VCTK.ipynb
    ├── train_list.txt
    └── val_list.txt
├── Demo
    ├── VCTK-corpus
    │   ├── p228
    │   │   └── p228_023.wav
    │   ├── p230
    │   │   └── p230_023.wav
    │   ├── p233
    │   │   └── p233_023.wav
    │   ├── p236
    │   │   └── p236_023.wav
    │   ├── p243
    │   │   └── p243_023.wav
    │   ├── p244
    │   │   └── p244_023.wav
    │   ├── p254
    │   │   └── p254_023.wav
    │   ├── p258
    │   │   └── p258_023.wav
    │   ├── p259
    │   │   └── p259_023.wav
    │   └── p273
    │   │   └── p273_023.wav
    └── inference.ipynb
├── LICENSE
├── README.md
├── Utils
    ├── ASR
    │   ├── __init__.py
    │   ├── config.yml
    │   ├── epoch_00100.pth
    │   ├── layers.py
    │   └── models.py
    ├── JDC
    │   ├── __init__.py
    │   ├── bst.t7
    │   └── model.py
    └── __init__.py
├── losses.py
├── meldataset.py
├── models.py
├── optimizers.py
├── train.py
├── trainer.py
└── transforms.py


/Configs/config.yml:
--------------------------------------------------------------------------------
 1 | log_dir: "Models/VCTK20"
 2 | save_freq: 2
 3 | device: "cuda"
 4 | epochs: 150
 5 | batch_size: 5
 6 | pretrained_model: ""
 7 | load_only_params: false
 8 | fp16_run: true
 9 | 
10 | train_data: "Data/train_list.txt"
11 | val_data: "Data/val_list.txt"
12 | 
13 | F0_path: "Utils/JDC/bst.t7"
14 | ASR_config: "Utils/ASR/config.yml"
15 | ASR_path: "Utils/ASR/epoch_00100.pth"
16 | 
17 | preprocess_params:
18 |   sr: 24000
19 |   spect_params:
20 |     n_fft: 2048
21 |     win_length: 1200
22 |     hop_length: 300
23 | 
24 | model_params:
25 |   dim_in: 64
26 |   style_dim: 64
27 |   latent_dim: 16
28 |   num_domains: 20
29 |   max_conv_dim: 512
30 |   n_repeat: 4
31 |   w_hpf: 0
32 |   F0_channel: 256
33 | 
34 | loss_params:
35 |   g_loss:
36 |     lambda_sty: 1.
37 |     lambda_cyc: 5.
38 |     lambda_ds: 1.
39 |     lambda_norm: 1.
40 |     lambda_asr: 10.
41 |     lambda_f0: 5.
42 |     lambda_f0_sty: 0.1
43 |     lambda_adv: 2.
44 |     lambda_adv_cls: 0.5
45 |     norm_bias: 0.5
46 |   d_loss:
47 |     lambda_reg: 1.
48 |     lambda_adv_cls: 0.1
49 |     lambda_con_reg: 10.
50 |   
51 |   adv_cls_epoch: 50
52 |   con_reg_epoch: 30
53 | 
54 | optimizer_params:
55 |   lr: 0.0001
56 | 


--------------------------------------------------------------------------------
/Data/VCTK.ipynb:
--------------------------------------------------------------------------------
  1 | {
  2 |  "cells": [
  3 |   {
  4 |    "cell_type": "code",
  5 |    "execution_count": null,
  6 |    "id": "347ace04",
  7 |    "metadata": {},
  8 |    "outputs": [],
  9 |    "source": [
 10 |     "# VCTK Corpus Path\n",
 11 |     "__CORPUSPATH__ = \"/share/naplab/users/yl4579/data/VCTK-Corpus/VCTK-Corpus\"\n",
 12 |     "\n",
 13 |     "# output path\n",
 14 |     "__OUTPATH__ = \"./Data\""
 15 |    ]
 16 |   },
 17 |   {
 18 |    "cell_type": "code",
 19 |    "execution_count": null,
 20 |    "id": "4ce9eb2e",
 21 |    "metadata": {},
 22 |    "outputs": [],
 23 |    "source": [
 24 |     "import os\n",
 25 |     "from scipy.io import wavfile\n",
 26 |     "from pydub import AudioSegment\n",
 27 |     "\n",
 28 |     "from pydub import AudioSegment\n",
 29 |     "from pydub.silence import split_on_silence\n",
 30 |     "import os\n",
 31 |     "\n",
 32 |     "def split(sound):\n",
 33 |     "    dBFS = sound.dBFS\n",
 34 |     "    chunks = split_on_silence(sound,\n",
 35 |     "        min_silence_len = 100,\n",
 36 |     "        silence_thresh = dBFS-16,\n",
 37 |     "        keep_silence = 100\n",
 38 |     "    )\n",
 39 |     "    return chunks\n",
 40 |     "\n",
 41 |     "def combine(_src):\n",
 42 |     "    audio = AudioSegment.empty()\n",
 43 |     "    for i,filename in enumerate(os.listdir(_src)):\n",
 44 |     "        if filename.endswith('.wav'):\n",
 45 |     "            filename = os.path.join(_src, filename)\n",
 46 |     "            audio += AudioSegment.from_wav(filename)\n",
 47 |     "    return audio\n",
 48 |     "\n",
 49 |     "def save_chunks(chunks, directory):\n",
 50 |     "    if not os.path.exists(directory):\n",
 51 |     "        os.makedirs(directory)\n",
 52 |     "    counter = 0\n",
 53 |     "\n",
 54 |     "    target_length = 5 * 1000\n",
 55 |     "    output_chunks = [chunks[0]]\n",
 56 |     "    for chunk in chunks[1:]:\n",
 57 |     "        if len(output_chunks[-1]) < target_length:\n",
 58 |     "            output_chunks[-1] += chunk\n",
 59 |     "        else:\n",
 60 |     "            # if the last output chunk is longer than the target length,\n",
 61 |     "            # we can start a new one\n",
 62 |     "            output_chunks.append(chunk)\n",
 63 |     "\n",
 64 |     "    for chunk in output_chunks:\n",
 65 |     "        chunk = chunk.set_frame_rate(24000)\n",
 66 |     "        chunk = chunk.set_channels(1)\n",
 67 |     "        counter = counter + 1\n",
 68 |     "        chunk.export(os.path.join(directory, str(counter) + '.wav'), format=\"wav\")"
 69 |    ]
 70 |   },
 71 |   {
 72 |    "cell_type": "code",
 73 |    "execution_count": null,
 74 |    "id": "769a7f62",
 75 |    "metadata": {},
 76 |    "outputs": [],
 77 |    "source": [
 78 |     "# Source: http://speech.ee.ntu.edu.tw/~jjery2243542/resource/model/is18/en_speaker_used.txt\n",
 79 |     "# Source: https://github.com/jjery2243542/voice_conversion\n",
 80 |     "\n",
 81 |     "speakers = [225,228,229,230,231,233,236,239,240,244,226,227,232,243,254,256,258,259,270,273]"
 82 |    ]
 83 |   },
 84 |   {
 85 |    "cell_type": "code",
 86 |    "execution_count": null,
 87 |    "id": "9302fb6a",
 88 |    "metadata": {},
 89 |    "outputs": [],
 90 |    "source": [
 91 |     "# downsample to 24 kHz\n",
 92 |     "\n",
 93 |     "for p in speakers:\n",
 94 |     "    directory = __OUTPATH__ + '/p' + str(p)\n",
 95 |     "    if not os.path.exists(directory):\n",
 96 |     "        audio = combine(__CORPUSPATH__ + '/wav48/p' + str(p))\n",
 97 |     "        chunks = split(audio)\n",
 98 |     "        save_chunks(chunks, directory)"
 99 |    ]
100 |   },
101 |   {
102 |    "cell_type": "code",
103 |    "execution_count": null,
104 |    "id": "4b0ca022",
105 |    "metadata": {},
106 |    "outputs": [],
107 |    "source": [
108 |     "# get all speakers\n",
109 |     "\n",
110 |     "data_list = []\n",
111 |     "for path, subdirs, files in os.walk(__OUTPATH__):\n",
112 |     "    for name in files:\n",
113 |     "        if name.endswith(\".wav\"):\n",
114 |     "            speaker = int(path.split('/')[-1].replace('p', ''))\n",
115 |     "            if speaker in speakers:\n",
116 |     "                data_list.append({\"Path\": os.path.join(path, name), \"Speaker\": int(speakers.index(speaker)) + 1})\n",
117 |     "                \n",
118 |     "import pandas as pd\n",
119 |     "\n",
120 |     "data_list = pd.DataFrame(data_list)\n",
121 |     "data_list = data_list.sample(frac=1)\n",
122 |     "\n",
123 |     "import random\n",
124 |     "\n",
125 |     "split_idx = round(len(data_list) * 0.1)\n",
126 |     "\n",
127 |     "test_data = data_list[:split_idx]\n",
128 |     "train_data = data_list[split_idx:]"
129 |    ]
130 |   },
131 |   {
132 |    "cell_type": "code",
133 |    "execution_count": null,
134 |    "id": "88df2a45",
135 |    "metadata": {},
136 |    "outputs": [],
137 |    "source": [
138 |     "# write to file \n",
139 |     "\n",
140 |     "file_str = \"\"\n",
141 |     "for index, k in train_data.iterrows():\n",
142 |     "    file_str += k['Path'] + \"|\" +str(k['Speaker'] - 1)+ '\\n'\n",
143 |     "text_file = open(__OUTPATH__ + \"/train_list.txt\", \"w\")\n",
144 |     "text_file.write(file_str)\n",
145 |     "text_file.close()\n",
146 |     "\n",
147 |     "file_str = \"\"\n",
148 |     "for index, k in test_data.iterrows():\n",
149 |     "    file_str += k['Path'] + \"|\" + str(k['Speaker'] - 1) + '\\n'\n",
150 |     "text_file = open(__OUTPATH__ + \"/val_list.txt\", \"w\")\n",
151 |     "text_file.write(file_str)\n",
152 |     "text_file.close()"
153 |    ]
154 |   }
155 |  ],
156 |  "metadata": {
157 |   "kernelspec": {
158 |    "display_name": "Python 3",
159 |    "language": "python",
160 |    "name": "python3"
161 |   },
162 |   "language_info": {
163 |    "codemirror_mode": {
164 |     "name": "ipython",
165 |     "version": 3
166 |    },
167 |    "file_extension": ".py",
168 |    "mimetype": "text/x-python",
169 |    "name": "python",
170 |    "nbconvert_exporter": "python",
171 |    "pygments_lexer": "ipython3",
172 |    "version": "3.7.10"
173 |   }
174 |  },
175 |  "nbformat": 4,
176 |  "nbformat_minor": 5
177 | }
178 | 


--------------------------------------------------------------------------------
/Data/train_list.txt:
--------------------------------------------------------------------------------
   1 | ./Data/p225/22.wav|0
   2 | ./Data/p239/163.wav|7
   3 | ./Data/p227/144.wav|11
   4 | ./Data/p258/67.wav|16
   5 | ./Data/p259/74.wav|17
   6 | ./Data/p230/103.wav|3
   7 | ./Data/p225/7.wav|0
   8 | ./Data/p226/67.wav|10
   9 | ./Data/p228/27.wav|1
  10 | ./Data/p243/23.wav|13
  11 | ./Data/p228/3.wav|1
  12 | ./Data/p244/61.wav|9
  13 | ./Data/p230/178.wav|3
  14 | ./Data/p240/120.wav|8
  15 | ./Data/p228/154.wav|1
  16 | ./Data/p230/146.wav|3
  17 | ./Data/p240/149.wav|8
  18 | ./Data/p254/56.wav|14
  19 | ./Data/p240/81.wav|8
  20 | ./Data/p226/27.wav|10
  21 | ./Data/p256/59.wav|15
  22 | ./Data/p231/74.wav|4
  23 | ./Data/p231/9.wav|4
  24 | ./Data/p240/1.wav|8
  25 | ./Data/p236/135.wav|6
  26 | ./Data/p232/85.wav|12
  27 | ./Data/p230/69.wav|3
  28 | ./Data/p256/35.wav|15
  29 | ./Data/p239/6.wav|7
  30 | ./Data/p254/70.wav|14
  31 | ./Data/p244/135.wav|9
  32 | ./Data/p254/114.wav|14
  33 | ./Data/p236/117.wav|6
  34 | ./Data/p225/78.wav|0
  35 | ./Data/p236/66.wav|6
  36 | ./Data/p228/155.wav|1
  37 | ./Data/p239/83.wav|7
  38 | ./Data/p240/22.wav|8
  39 | ./Data/p225/2.wav|0
  40 | ./Data/p230/75.wav|3
  41 | ./Data/p239/17.wav|7
  42 | ./Data/p239/147.wav|7
  43 | ./Data/p273/125.wav|19
  44 | ./Data/p270/17.wav|18
  45 | ./Data/p233/74.wav|5
  46 | ./Data/p233/112.wav|5
  47 | ./Data/p228/1.wav|1
  48 | ./Data/p258/125.wav|16
  49 | ./Data/p231/56.wav|4
  50 | ./Data/p227/101.wav|11
  51 | ./Data/p232/108.wav|12
  52 | ./Data/p239/82.wav|7
  53 | ./Data/p270/160.wav|18
  54 | ./Data/p227/58.wav|11
  55 | ./Data/p233/111.wav|5
  56 | ./Data/p259/146.wav|17
  57 | ./Data/p230/77.wav|3
  58 | ./Data/p256/104.wav|15
  59 | ./Data/p228/140.wav|1
  60 | ./Data/p231/143.wav|4
  61 | ./Data/p270/20.wav|18
  62 | ./Data/p225/62.wav|0
  63 | ./Data/p229/11.wav|2
  64 | ./Data/p259/66.wav|17
  65 | ./Data/p239/53.wav|7
  66 | ./Data/p239/23.wav|7
  67 | ./Data/p240/115.wav|8
  68 | ./Data/p233/41.wav|5
  69 | ./Data/p270/61.wav|18
  70 | ./Data/p232/50.wav|12
  71 | ./Data/p239/56.wav|7
  72 | ./Data/p244/117.wav|9
  73 | ./Data/p233/6.wav|5
  74 | ./Data/p227/31.wav|11
  75 | ./Data/p231/134.wav|4
  76 | ./Data/p243/12.wav|13
  77 | ./Data/p226/14.wav|10
  78 | ./Data/p240/45.wav|8
  79 | ./Data/p231/91.wav|4
  80 | ./Data/p259/163.wav|17
  81 | ./Data/p236/41.wav|6
  82 | ./Data/p231/66.wav|4
  83 | ./Data/p233/122.wav|5
  84 | ./Data/p244/8.wav|9
  85 | ./Data/p232/41.wav|12
  86 | ./Data/p232/101.wav|12
  87 | ./Data/p273/70.wav|19
  88 | ./Data/p270/168.wav|18
  89 | ./Data/p226/80.wav|10
  90 | ./Data/p270/178.wav|18
  91 | ./Data/p225/39.wav|0
  92 | ./Data/p258/29.wav|16
  93 | ./Data/p231/46.wav|4
  94 | ./Data/p244/41.wav|9
  95 | ./Data/p227/115.wav|11
  96 | ./Data/p228/14.wav|1
  97 | ./Data/p239/116.wav|7
  98 | ./Data/p259/13.wav|17
  99 | ./Data/p254/51.wav|14
 100 | ./Data/p256/36.wav|15
 101 | ./Data/p254/108.wav|14
 102 | ./Data/p226/46.wav|10
 103 | ./Data/p258/39.wav|16
 104 | ./Data/p273/106.wav|19
 105 | ./Data/p228/104.wav|1
 106 | ./Data/p256/1.wav|15
 107 | ./Data/p258/109.wav|16
 108 | ./Data/p259/123.wav|17
 109 | ./Data/p258/99.wav|16
 110 | ./Data/p256/61.wav|15
 111 | ./Data/p231/17.wav|4
 112 | ./Data/p227/74.wav|11
 113 | ./Data/p256/21.wav|15
 114 | ./Data/p226/1.wav|10
 115 | ./Data/p231/129.wav|4
 116 | ./Data/p231/79.wav|4
 117 | ./Data/p226/74.wav|10
 118 | ./Data/p233/69.wav|5
 119 | ./Data/p227/44.wav|11
 120 | ./Data/p239/141.wav|7
 121 | ./Data/p228/28.wav|1
 122 | ./Data/p239/101.wav|7
 123 | ./Data/p258/106.wav|16
 124 | ./Data/p236/49.wav|6
 125 | ./Data/p230/46.wav|3
 126 | ./Data/p244/118.wav|9
 127 | ./Data/p227/25.wav|11
 128 | ./Data/p243/25.wav|13
 129 | ./Data/p270/96.wav|18
 130 | ./Data/p228/130.wav|1
 131 | ./Data/p230/62.wav|3
 132 | ./Data/p227/114.wav|11
 133 | ./Data/p228/116.wav|1
 134 | ./Data/p233/40.wav|5
 135 | ./Data/p230/147.wav|3
 136 | ./Data/p240/71.wav|8
 137 | ./Data/p233/29.wav|5
 138 | ./Data/p230/13.wav|3
 139 | ./Data/p239/63.wav|7
 140 | ./Data/p228/19.wav|1
 141 | ./Data/p233/63.wav|5
 142 | ./Data/p227/50.wav|11
 143 | ./Data/p270/6.wav|18
 144 | ./Data/p228/106.wav|1
 145 | ./Data/p236/86.wav|6
 146 | ./Data/p240/129.wav|8
 147 | ./Data/p273/21.wav|19
 148 | ./Data/p256/83.wav|15
 149 | ./Data/p240/85.wav|8
 150 | ./Data/p258/43.wav|16
 151 | ./Data/p273/41.wav|19
 152 | ./Data/p270/192.wav|18
 153 | ./Data/p230/134.wav|3
 154 | ./Data/p228/44.wav|1
 155 | ./Data/p231/102.wav|4
 156 | ./Data/p270/73.wav|18
 157 | ./Data/p239/153.wav|7
 158 | ./Data/p270/36.wav|18
 159 | ./Data/p273/145.wav|19
 160 | ./Data/p228/107.wav|1
 161 | ./Data/p244/12.wav|9
 162 | ./Data/p270/181.wav|18
 163 | ./Data/p231/35.wav|4
 164 | ./Data/p233/132.wav|5
 165 | ./Data/p226/19.wav|10
 166 | ./Data/p239/91.wav|7
 167 | ./Data/p225/66.wav|0
 168 | ./Data/p229/40.wav|2
 169 | ./Data/p227/48.wav|11
 170 | ./Data/p225/44.wav|0
 171 | ./Data/p229/108.wav|2
 172 | ./Data/p227/47.wav|11
 173 | ./Data/p270/78.wav|18
 174 | ./Data/p259/77.wav|17
 175 | ./Data/p239/51.wav|7
 176 | ./Data/p230/58.wav|3
 177 | ./Data/p233/81.wav|5
 178 | ./Data/p230/150.wav|3
 179 | ./Data/p227/141.wav|11
 180 | ./Data/p243/121.wav|13
 181 | ./Data/p244/11.wav|9
 182 | ./Data/p270/63.wav|18
 183 | ./Data/p236/81.wav|6
 184 | ./Data/p229/89.wav|2
 185 | ./Data/p231/83.wav|4
 186 | ./Data/p233/33.wav|5
 187 | ./Data/p227/107.wav|11
 188 | ./Data/p228/48.wav|1
 189 | ./Data/p259/44.wav|17
 190 | ./Data/p228/131.wav|1
 191 | ./Data/p227/20.wav|11
 192 | ./Data/p256/98.wav|15
 193 | ./Data/p273/45.wav|19
 194 | ./Data/p239/137.wav|7
 195 | ./Data/p232/93.wav|12
 196 | ./Data/p239/38.wav|7
 197 | ./Data/p243/161.wav|13
 198 | ./Data/p258/3.wav|16
 199 | ./Data/p273/132.wav|19
 200 | ./Data/p230/15.wav|3
 201 | ./Data/p259/155.wav|17
 202 | ./Data/p256/125.wav|15
 203 | ./Data/p256/7.wav|15
 204 | ./Data/p231/58.wav|4
 205 | ./Data/p256/75.wav|15
 206 | ./Data/p236/116.wav|6
 207 | ./Data/p233/10.wav|5
 208 | ./Data/p270/39.wav|18
 209 | ./Data/p254/48.wav|14
 210 | ./Data/p270/121.wav|18
 211 | ./Data/p240/139.wav|8
 212 | ./Data/p240/121.wav|8
 213 | ./Data/p244/13.wav|9
 214 | ./Data/p243/148.wav|13
 215 | ./Data/p240/125.wav|8
 216 | ./Data/p259/14.wav|17
 217 | ./Data/p273/139.wav|19
 218 | ./Data/p233/123.wav|5
 219 | ./Data/p225/34.wav|0
 220 | ./Data/p244/10.wav|9
 221 | ./Data/p258/4.wav|16
 222 | ./Data/p236/111.wav|6
 223 | ./Data/p259/59.wav|17
 224 | ./Data/p258/85.wav|16
 225 | ./Data/p227/73.wav|11
 226 | ./Data/p273/107.wav|19
 227 | ./Data/p231/94.wav|4
 228 | ./Data/p231/34.wav|4
 229 | ./Data/p270/200.wav|18
 230 | ./Data/p273/74.wav|19
 231 | ./Data/p232/45.wav|12
 232 | ./Data/p227/37.wav|11
 233 | ./Data/p256/101.wav|15
 234 | ./Data/p233/64.wav|5
 235 | ./Data/p228/52.wav|1
 236 | ./Data/p254/104.wav|14
 237 | ./Data/p236/103.wav|6
 238 | ./Data/p233/73.wav|5
 239 | ./Data/p243/146.wav|13
 240 | ./Data/p258/60.wav|16
 241 | ./Data/p254/8.wav|14
 242 | ./Data/p226/64.wav|10
 243 | ./Data/p243/63.wav|13
 244 | ./Data/p226/70.wav|10
 245 | ./Data/p233/37.wav|5
 246 | ./Data/p254/78.wav|14
 247 | ./Data/p227/123.wav|11
 248 | ./Data/p240/55.wav|8
 249 | ./Data/p229/126.wav|2
 250 | ./Data/p254/54.wav|14
 251 | ./Data/p243/163.wav|13
 252 | ./Data/p273/79.wav|19
 253 | ./Data/p230/9.wav|3
 254 | ./Data/p270/50.wav|18
 255 | ./Data/p243/64.wav|13
 256 | ./Data/p229/100.wav|2
 257 | ./Data/p240/100.wav|8
 258 | ./Data/p239/139.wav|7
 259 | ./Data/p236/65.wav|6
 260 | ./Data/p243/97.wav|13
 261 | ./Data/p258/37.wav|16
 262 | ./Data/p233/9.wav|5
 263 | ./Data/p243/10.wav|13
 264 | ./Data/p244/53.wav|9
 265 | ./Data/p259/162.wav|17
 266 | ./Data/p236/131.wav|6
 267 | ./Data/p227/134.wav|11
 268 | ./Data/p228/5.wav|1
 269 | ./Data/p273/18.wav|19
 270 | ./Data/p243/115.wav|13
 271 | ./Data/p256/113.wav|15
 272 | ./Data/p243/103.wav|13
 273 | ./Data/p273/133.wav|19
 274 | ./Data/p244/90.wav|9
 275 | ./Data/p258/45.wav|16
 276 | ./Data/p229/16.wav|2
 277 | ./Data/p244/25.wav|9
 278 | ./Data/p225/95.wav|0
 279 | ./Data/p230/18.wav|3
 280 | ./Data/p270/196.wav|18
 281 | ./Data/p229/58.wav|2
 282 | ./Data/p239/125.wav|7
 283 | ./Data/p225/27.wav|0
 284 | ./Data/p239/13.wav|7
 285 | ./Data/p259/121.wav|17
 286 | ./Data/p240/12.wav|8
 287 | ./Data/p270/40.wav|18
 288 | ./Data/p258/17.wav|16
 289 | ./Data/p270/123.wav|18
 290 | ./Data/p258/93.wav|16
 291 | ./Data/p229/43.wav|2
 292 | ./Data/p243/152.wav|13
 293 | ./Data/p236/11.wav|6
 294 | ./Data/p232/38.wav|12
 295 | ./Data/p225/9.wav|0
 296 | ./Data/p270/128.wav|18
 297 | ./Data/p258/22.wav|16
 298 | ./Data/p227/113.wav|11
 299 | ./Data/p228/128.wav|1
 300 | ./Data/p228/56.wav|1
 301 | ./Data/p239/19.wav|7
 302 | ./Data/p273/134.wav|19
 303 | ./Data/p231/144.wav|4
 304 | ./Data/p231/16.wav|4
 305 | ./Data/p259/141.wav|17
 306 | ./Data/p227/9.wav|11
 307 | ./Data/p273/114.wav|19
 308 | ./Data/p225/94.wav|0
 309 | ./Data/p273/42.wav|19
 310 | ./Data/p236/72.wav|6
 311 | ./Data/p240/58.wav|8
 312 | ./Data/p258/78.wav|16
 313 | ./Data/p227/129.wav|11
 314 | ./Data/p254/9.wav|14
 315 | ./Data/p226/43.wav|10
 316 | ./Data/p228/103.wav|1
 317 | ./Data/p232/114.wav|12
 318 | ./Data/p254/80.wav|14
 319 | ./Data/p240/144.wav|8
 320 | ./Data/p227/55.wav|11
 321 | ./Data/p254/2.wav|14
 322 | ./Data/p273/101.wav|19
 323 | ./Data/p243/67.wav|13
 324 | ./Data/p227/94.wav|11
 325 | ./Data/p227/121.wav|11
 326 | ./Data/p259/153.wav|17
 327 | ./Data/p258/40.wav|16
 328 | ./Data/p239/32.wav|7
 329 | ./Data/p270/83.wav|18
 330 | ./Data/p226/103.wav|10
 331 | ./Data/p258/18.wav|16
 332 | ./Data/p243/155.wav|13
 333 | ./Data/p229/117.wav|2
 334 | ./Data/p231/127.wav|4
 335 | ./Data/p256/30.wav|15
 336 | ./Data/p240/90.wav|8
 337 | ./Data/p254/133.wav|14
 338 | ./Data/p240/51.wav|8
 339 | ./Data/p239/105.wav|7
 340 | ./Data/p226/85.wav|10
 341 | ./Data/p254/31.wav|14
 342 | ./Data/p258/57.wav|16
 343 | ./Data/p230/95.wav|3
 344 | ./Data/p226/52.wav|10
 345 | ./Data/p258/79.wav|16
 346 | ./Data/p273/49.wav|19
 347 | ./Data/p259/82.wav|17
 348 | ./Data/p227/126.wav|11
 349 | ./Data/p243/158.wav|13
 350 | ./Data/p273/130.wav|19
 351 | ./Data/p243/7.wav|13
 352 | ./Data/p228/137.wav|1
 353 | ./Data/p233/103.wav|5
 354 | ./Data/p254/37.wav|14
 355 | ./Data/p240/39.wav|8
 356 | ./Data/p225/28.wav|0
 357 | ./Data/p227/139.wav|11
 358 | ./Data/p244/130.wav|9
 359 | ./Data/p243/22.wav|13
 360 | ./Data/p228/69.wav|1
 361 | ./Data/p231/64.wav|4
 362 | ./Data/p233/129.wav|5
 363 | ./Data/p232/68.wav|12
 364 | ./Data/p231/87.wav|4
 365 | ./Data/p240/83.wav|8
 366 | ./Data/p232/55.wav|12
 367 | ./Data/p259/54.wav|17
 368 | ./Data/p270/125.wav|18
 369 | ./Data/p239/169.wav|7
 370 | ./Data/p229/39.wav|2
 371 | ./Data/p273/110.wav|19
 372 | ./Data/p233/93.wav|5
 373 | ./Data/p225/79.wav|0
 374 | ./Data/p230/140.wav|3
 375 | ./Data/p228/36.wav|1
 376 | ./Data/p230/4.wav|3
 377 | ./Data/p259/88.wav|17
 378 | ./Data/p243/86.wav|13
 379 | ./Data/p227/90.wav|11
 380 | ./Data/p254/83.wav|14
 381 | ./Data/p240/150.wav|8
 382 | ./Data/p232/118.wav|12
 383 | ./Data/p270/35.wav|18
 384 | ./Data/p231/126.wav|4
 385 | ./Data/p239/59.wav|7
 386 | ./Data/p243/47.wav|13
 387 | ./Data/p254/105.wav|14
 388 | ./Data/p258/72.wav|16
 389 | ./Data/p228/72.wav|1
 390 | ./Data/p270/136.wav|18
 391 | ./Data/p230/51.wav|3
 392 | ./Data/p227/67.wav|11
 393 | ./Data/p259/151.wav|17
 394 | ./Data/p232/66.wav|12
 395 | ./Data/p254/40.wav|14
 396 | ./Data/p273/12.wav|19
 397 | ./Data/p229/130.wav|2
 398 | ./Data/p270/156.wav|18
 399 | ./Data/p230/177.wav|3
 400 | ./Data/p270/169.wav|18
 401 | ./Data/p258/1.wav|16
 402 | ./Data/p229/103.wav|2
 403 | ./Data/p270/127.wav|18
 404 | ./Data/p226/72.wav|10
 405 | ./Data/p229/99.wav|2
 406 | ./Data/p232/8.wav|12
 407 | ./Data/p236/1.wav|6
 408 | ./Data/p230/85.wav|3
 409 | ./Data/p236/99.wav|6
 410 | ./Data/p231/139.wav|4
 411 | ./Data/p256/67.wav|15
 412 | ./Data/p240/38.wav|8
 413 | ./Data/p233/16.wav|5
 414 | ./Data/p243/13.wav|13
 415 | ./Data/p227/86.wav|11
 416 | ./Data/p233/110.wav|5
 417 | ./Data/p243/77.wav|13
 418 | ./Data/p227/77.wav|11
 419 | ./Data/p230/7.wav|3
 420 | ./Data/p270/175.wav|18
 421 | ./Data/p254/38.wav|14
 422 | ./Data/p227/71.wav|11
 423 | ./Data/p229/104.wav|2
 424 | ./Data/p231/101.wav|4
 425 | ./Data/p229/105.wav|2
 426 | ./Data/p225/49.wav|0
 427 | ./Data/p230/137.wav|3
 428 | ./Data/p226/42.wav|10
 429 | ./Data/p233/92.wav|5
 430 | ./Data/p243/58.wav|13
 431 | ./Data/p239/45.wav|7
 432 | ./Data/p233/135.wav|5
 433 | ./Data/p244/89.wav|9
 434 | ./Data/p243/166.wav|13
 435 | ./Data/p240/59.wav|8
 436 | ./Data/p254/86.wav|14
 437 | ./Data/p243/60.wav|13
 438 | ./Data/p227/19.wav|11
 439 | ./Data/p231/45.wav|4
 440 | ./Data/p227/140.wav|11
 441 | ./Data/p236/129.wav|6
 442 | ./Data/p240/67.wav|8
 443 | ./Data/p227/61.wav|11
 444 | ./Data/p228/77.wav|1
 445 | ./Data/p236/52.wav|6
 446 | ./Data/p258/33.wav|16
 447 | ./Data/p244/104.wav|9
 448 | ./Data/p259/84.wav|17
 449 | ./Data/p236/127.wav|6
 450 | ./Data/p228/150.wav|1
 451 | ./Data/p233/85.wav|5
 452 | ./Data/p270/147.wav|18
 453 | ./Data/p229/83.wav|2
 454 | ./Data/p226/68.wav|10
 455 | ./Data/p229/94.wav|2
 456 | ./Data/p270/46.wav|18
 457 | ./Data/p258/129.wav|16
 458 | ./Data/p270/191.wav|18
 459 | ./Data/p227/106.wav|11
 460 | ./Data/p239/136.wav|7
 461 | ./Data/p239/14.wav|7
 462 | ./Data/p239/71.wav|7
 463 | ./Data/p232/74.wav|12
 464 | ./Data/p225/75.wav|0
 465 | ./Data/p244/143.wav|9
 466 | ./Data/p259/173.wav|17
 467 | ./Data/p243/140.wav|13
 468 | ./Data/p273/48.wav|19
 469 | ./Data/p230/111.wav|3
 470 | ./Data/p240/94.wav|8
 471 | ./Data/p258/20.wav|16
 472 | ./Data/p227/52.wav|11
 473 | ./Data/p244/4.wav|9
 474 | ./Data/p227/109.wav|11
 475 | ./Data/p230/55.wav|3
 476 | ./Data/p232/92.wav|12
 477 | ./Data/p240/75.wav|8
 478 | ./Data/p229/82.wav|2
 479 | ./Data/p270/103.wav|18
 480 | ./Data/p254/87.wav|14
 481 | ./Data/p259/38.wav|17
 482 | ./Data/p240/147.wav|8
 483 | ./Data/p227/111.wav|11
 484 | ./Data/p228/2.wav|1
 485 | ./Data/p230/82.wav|3
 486 | ./Data/p239/33.wav|7
 487 | ./Data/p259/65.wav|17
 488 | ./Data/p273/102.wav|19
 489 | ./Data/p227/116.wav|11
 490 | ./Data/p258/61.wav|16
 491 | ./Data/p228/68.wav|1
 492 | ./Data/p244/116.wav|9
 493 | ./Data/p240/9.wav|8
 494 | ./Data/p273/64.wav|19
 495 | ./Data/p273/9.wav|19
 496 | ./Data/p230/8.wav|3
 497 | ./Data/p230/172.wav|3
 498 | ./Data/p243/32.wav|13
 499 | ./Data/p258/117.wav|16
 500 | ./Data/p236/43.wav|6
 501 | ./Data/p243/29.wav|13
 502 | ./Data/p231/86.wav|4
 503 | ./Data/p231/6.wav|4
 504 | ./Data/p236/166.wav|6
 505 | ./Data/p270/174.wav|18
 506 | ./Data/p229/123.wav|2
 507 | ./Data/p243/132.wav|13
 508 | ./Data/p228/91.wav|1
 509 | ./Data/p273/100.wav|19
 510 | ./Data/p243/61.wav|13
 511 | ./Data/p233/14.wav|5
 512 | ./Data/p256/5.wav|15
 513 | ./Data/p228/135.wav|1
 514 | ./Data/p254/21.wav|14
 515 | ./Data/p230/96.wav|3
 516 | ./Data/p240/142.wav|8
 517 | ./Data/p259/63.wav|17
 518 | ./Data/p243/37.wav|13
 519 | ./Data/p228/136.wav|1
 520 | ./Data/p254/126.wav|14
 521 | ./Data/p225/51.wav|0
 522 | ./Data/p258/9.wav|16
 523 | ./Data/p270/85.wav|18
 524 | ./Data/p228/149.wav|1
 525 | ./Data/p236/152.wav|6
 526 | ./Data/p259/124.wav|17
 527 | ./Data/p244/1.wav|9
 528 | ./Data/p259/104.wav|17
 529 | ./Data/p227/64.wav|11
 530 | ./Data/p230/70.wav|3
 531 | ./Data/p256/122.wav|15
 532 | ./Data/p258/30.wav|16
 533 | ./Data/p244/54.wav|9
 534 | ./Data/p270/198.wav|18
 535 | ./Data/p258/15.wav|16
 536 | ./Data/p254/52.wav|14
 537 | ./Data/p228/85.wav|1
 538 | ./Data/p230/1.wav|3
 539 | ./Data/p230/71.wav|3
 540 | ./Data/p259/147.wav|17
 541 | ./Data/p243/68.wav|13
 542 | ./Data/p226/79.wav|10
 543 | ./Data/p243/123.wav|13
 544 | ./Data/p229/85.wav|2
 545 | ./Data/p270/5.wav|18
 546 | ./Data/p226/12.wav|10
 547 | ./Data/p231/82.wav|4
 548 | ./Data/p230/120.wav|3
 549 | ./Data/p225/31.wav|0
 550 | ./Data/p236/130.wav|6
 551 | ./Data/p239/111.wav|7
 552 | ./Data/p230/60.wav|3
 553 | ./Data/p232/121.wav|12
 554 | ./Data/p259/27.wav|17
 555 | ./Data/p228/65.wav|1
 556 | ./Data/p231/92.wav|4
 557 | ./Data/p236/160.wav|6
 558 | ./Data/p258/145.wav|16
 559 | ./Data/p231/20.wav|4
 560 | ./Data/p226/47.wav|10
 561 | ./Data/p258/110.wav|16
 562 | ./Data/p231/93.wav|4
 563 | ./Data/p270/30.wav|18
 564 | ./Data/p227/97.wav|11
 565 | ./Data/p231/31.wav|4
 566 | ./Data/p273/55.wav|19
 567 | ./Data/p239/12.wav|7
 568 | ./Data/p240/63.wav|8
 569 | ./Data/p254/57.wav|14
 570 | ./Data/p244/35.wav|9
 571 | ./Data/p239/127.wav|7
 572 | ./Data/p226/130.wav|10
 573 | ./Data/p225/83.wav|0
 574 | ./Data/p259/56.wav|17
 575 | ./Data/p273/85.wav|19
 576 | ./Data/p244/129.wav|9
 577 | ./Data/p273/83.wav|19
 578 | ./Data/p270/45.wav|18
 579 | ./Data/p273/23.wav|19
 580 | ./Data/p233/15.wav|5
 581 | ./Data/p256/34.wav|15
 582 | ./Data/p273/38.wav|19
 583 | ./Data/p244/73.wav|9
 584 | ./Data/p243/43.wav|13
 585 | ./Data/p270/26.wav|18
 586 | ./Data/p239/87.wav|7
 587 | ./Data/p233/120.wav|5
 588 | ./Data/p236/14.wav|6
 589 | ./Data/p227/5.wav|11
 590 | ./Data/p258/104.wav|16
 591 | ./Data/p227/45.wav|11
 592 | ./Data/p229/35.wav|2
 593 | ./Data/p273/36.wav|19
 594 | ./Data/p240/82.wav|8
 595 | ./Data/p254/20.wav|14
 596 | ./Data/p232/128.wav|12
 597 | ./Data/p254/47.wav|14
 598 | ./Data/p270/102.wav|18
 599 | ./Data/p230/41.wav|3
 600 | ./Data/p225/23.wav|0
 601 | ./Data/p258/38.wav|16
 602 | ./Data/p233/137.wav|5
 603 | ./Data/p254/94.wav|14
 604 | ./Data/p244/122.wav|9
 605 | ./Data/p229/51.wav|2
 606 | ./Data/p244/96.wav|9
 607 | ./Data/p273/119.wav|19
 608 | ./Data/p227/80.wav|11
 609 | ./Data/p225/1.wav|0
 610 | ./Data/p244/80.wav|9
 611 | ./Data/p233/108.wav|5
 612 | ./Data/p259/119.wav|17
 613 | ./Data/p226/8.wav|10
 614 | ./Data/p228/60.wav|1
 615 | ./Data/p233/71.wav|5
 616 | ./Data/p243/168.wav|13
 617 | ./Data/p226/136.wav|10
 618 | ./Data/p236/110.wav|6
 619 | ./Data/p228/23.wav|1
 620 | ./Data/p244/137.wav|9
 621 | ./Data/p240/33.wav|8
 622 | ./Data/p256/13.wav|15
 623 | ./Data/p243/6.wav|13
 624 | ./Data/p227/30.wav|11
 625 | ./Data/p244/28.wav|9
 626 | ./Data/p228/24.wav|1
 627 | ./Data/p243/147.wav|13
 628 | ./Data/p231/39.wav|4
 629 | ./Data/p254/93.wav|14
 630 | ./Data/p256/29.wav|15
 631 | ./Data/p258/119.wav|16
 632 | ./Data/p240/69.wav|8
 633 | ./Data/p232/102.wav|12
 634 | ./Data/p233/101.wav|5
 635 | ./Data/p270/29.wav|18
 636 | ./Data/p233/47.wav|5
 637 | ./Data/p259/17.wav|17
 638 | ./Data/p228/94.wav|1
 639 | ./Data/p231/21.wav|4
 640 | ./Data/p230/143.wav|3
 641 | ./Data/p270/204.wav|18
 642 | ./Data/p229/71.wav|2
 643 | ./Data/p232/13.wav|12
 644 | ./Data/p227/127.wav|11
 645 | ./Data/p258/105.wav|16
 646 | ./Data/p227/112.wav|11
 647 | ./Data/p270/59.wav|18
 648 | ./Data/p232/47.wav|12
 649 | ./Data/p236/112.wav|6
 650 | ./Data/p273/115.wav|19
 651 | ./Data/p236/20.wav|6
 652 | ./Data/p258/115.wav|16
 653 | ./Data/p256/24.wav|15
 654 | ./Data/p273/76.wav|19
 655 | ./Data/p231/3.wav|4
 656 | ./Data/p225/56.wav|0
 657 | ./Data/p259/150.wav|17
 658 | ./Data/p227/70.wav|11
 659 | ./Data/p230/2.wav|3
 660 | ./Data/p226/22.wav|10
 661 | ./Data/p243/127.wav|13
 662 | ./Data/p258/31.wav|16
 663 | ./Data/p233/89.wav|5
 664 | ./Data/p259/64.wav|17
 665 | ./Data/p259/96.wav|17
 666 | ./Data/p227/57.wav|11
 667 | ./Data/p232/132.wav|12
 668 | ./Data/p236/46.wav|6
 669 | ./Data/p232/70.wav|12
 670 | ./Data/p273/138.wav|19
 671 | ./Data/p244/99.wav|9
 672 | ./Data/p240/18.wav|8
 673 | ./Data/p243/145.wav|13
 674 | ./Data/p230/125.wav|3
 675 | ./Data/p243/49.wav|13
 676 | ./Data/p256/71.wav|15
 677 | ./Data/p258/133.wav|16
 678 | ./Data/p236/50.wav|6
 679 | ./Data/p270/122.wav|18
 680 | ./Data/p230/25.wav|3
 681 | ./Data/p236/124.wav|6
 682 | ./Data/p273/35.wav|19
 683 | ./Data/p258/98.wav|16
 684 | ./Data/p270/51.wav|18
 685 | ./Data/p229/121.wav|2
 686 | ./Data/p270/15.wav|18
 687 | ./Data/p270/193.wav|18
 688 | ./Data/p239/138.wav|7
 689 | ./Data/p273/108.wav|19
 690 | ./Data/p254/139.wav|14
 691 | ./Data/p256/23.wav|15
 692 | ./Data/p243/84.wav|13
 693 | ./Data/p273/93.wav|19
 694 | ./Data/p240/21.wav|8
 695 | ./Data/p240/109.wav|8
 696 | ./Data/p230/76.wav|3
 697 | ./Data/p232/61.wav|12
 698 | ./Data/p233/48.wav|5
 699 | ./Data/p233/133.wav|5
 700 | ./Data/p239/28.wav|7
 701 | ./Data/p230/149.wav|3
 702 | ./Data/p240/46.wav|8
 703 | ./Data/p243/74.wav|13
 704 | ./Data/p256/88.wav|15
 705 | ./Data/p228/61.wav|1
 706 | ./Data/p236/87.wav|6
 707 | ./Data/p236/2.wav|6
 708 | ./Data/p239/159.wav|7
 709 | ./Data/p231/44.wav|4
 710 | ./Data/p236/161.wav|6
 711 | ./Data/p256/19.wav|15
 712 | ./Data/p258/5.wav|16
 713 | ./Data/p243/83.wav|13
 714 | ./Data/p228/30.wav|1
 715 | ./Data/p226/65.wav|10
 716 | ./Data/p258/127.wav|16
 717 | ./Data/p254/60.wav|14
 718 | ./Data/p273/97.wav|19
 719 | ./Data/p228/50.wav|1
 720 | ./Data/p243/135.wav|13
 721 | ./Data/p228/111.wav|1
 722 | ./Data/p229/7.wav|2
 723 | ./Data/p229/3.wav|2
 724 | ./Data/p258/11.wav|16
 725 | ./Data/p258/6.wav|16
 726 | ./Data/p259/148.wav|17
 727 | ./Data/p232/30.wav|12
 728 | ./Data/p256/70.wav|15
 729 | ./Data/p259/160.wav|17
 730 | ./Data/p239/113.wav|7
 731 | ./Data/p229/109.wav|2
 732 | ./Data/p231/29.wav|4
 733 | ./Data/p258/25.wav|16
 734 | ./Data/p239/148.wav|7
 735 | ./Data/p239/78.wav|7
 736 | ./Data/p239/107.wav|7
 737 | ./Data/p239/99.wav|7
 738 | ./Data/p259/32.wav|17
 739 | ./Data/p239/11.wav|7
 740 | ./Data/p226/139.wav|10
 741 | ./Data/p229/88.wav|2
 742 | ./Data/p239/9.wav|7
 743 | ./Data/p229/26.wav|2
 744 | ./Data/p229/128.wav|2
 745 | ./Data/p244/119.wav|9
 746 | ./Data/p259/76.wav|17
 747 | ./Data/p239/129.wav|7
 748 | ./Data/p256/115.wav|15
 749 | ./Data/p230/102.wav|3
 750 | ./Data/p236/42.wav|6
 751 | ./Data/p225/16.wav|0
 752 | ./Data/p240/140.wav|8
 753 | ./Data/p226/36.wav|10
 754 | ./Data/p226/78.wav|10
 755 | ./Data/p225/37.wav|0
 756 | ./Data/p256/51.wav|15
 757 | ./Data/p254/112.wav|14
 758 | ./Data/p236/24.wav|6
 759 | ./Data/p228/164.wav|1
 760 | ./Data/p225/63.wav|0
 761 | ./Data/p259/25.wav|17
 762 | ./Data/p226/133.wav|10
 763 | ./Data/p244/107.wav|9
 764 | ./Data/p270/32.wav|18
 765 | ./Data/p270/56.wav|18
 766 | ./Data/p226/62.wav|10
 767 | ./Data/p228/95.wav|1
 768 | ./Data/p259/112.wav|17
 769 | ./Data/p229/114.wav|2
 770 | ./Data/p273/16.wav|19
 771 | ./Data/p236/60.wav|6
 772 | ./Data/p256/128.wav|15
 773 | ./Data/p273/144.wav|19
 774 | ./Data/p236/142.wav|6
 775 | ./Data/p231/130.wav|4
 776 | ./Data/p258/7.wav|16
 777 | ./Data/p225/96.wav|0
 778 | ./Data/p225/91.wav|0
 779 | ./Data/p232/115.wav|12
 780 | ./Data/p270/157.wav|18
 781 | ./Data/p273/104.wav|19
 782 | ./Data/p233/136.wav|5
 783 | ./Data/p240/78.wav|8
 784 | ./Data/p243/17.wav|13
 785 | ./Data/p240/62.wav|8
 786 | ./Data/p243/48.wav|13
 787 | ./Data/p232/29.wav|12
 788 | ./Data/p244/42.wav|9
 789 | ./Data/p259/93.wav|17
 790 | ./Data/p240/136.wav|8
 791 | ./Data/p226/117.wav|10
 792 | ./Data/p239/131.wav|7
 793 | ./Data/p270/54.wav|18
 794 | ./Data/p228/98.wav|1
 795 | ./Data/p270/166.wav|18
 796 | ./Data/p240/145.wav|8
 797 | ./Data/p270/14.wav|18
 798 | ./Data/p240/43.wav|8
 799 | ./Data/p258/107.wav|16
 800 | ./Data/p270/167.wav|18
 801 | ./Data/p259/62.wav|17
 802 | ./Data/p231/65.wav|4
 803 | ./Data/p240/5.wav|8
 804 | ./Data/p230/50.wav|3
 805 | ./Data/p256/3.wav|15
 806 | ./Data/p231/27.wav|4
 807 | ./Data/p229/27.wav|2
 808 | ./Data/p240/96.wav|8
 809 | ./Data/p225/82.wav|0
 810 | ./Data/p236/125.wav|6
 811 | ./Data/p254/71.wav|14
 812 | ./Data/p244/138.wav|9
 813 | ./Data/p254/89.wav|14
 814 | ./Data/p236/91.wav|6
 815 | ./Data/p244/38.wav|9
 816 | ./Data/p232/116.wav|12
 817 | ./Data/p270/11.wav|18
 818 | ./Data/p236/162.wav|6
 819 | ./Data/p228/127.wav|1
 820 | ./Data/p227/96.wav|11
 821 | ./Data/p226/98.wav|10
 822 | ./Data/p270/155.wav|18
 823 | ./Data/p236/143.wav|6
 824 | ./Data/p254/77.wav|14
 825 | ./Data/p273/26.wav|19
 826 | ./Data/p270/1.wav|18
 827 | ./Data/p273/51.wav|19
 828 | ./Data/p243/21.wav|13
 829 | ./Data/p231/68.wav|4
 830 | ./Data/p230/169.wav|3
 831 | ./Data/p226/56.wav|10
 832 | ./Data/p233/79.wav|5
 833 | ./Data/p273/58.wav|19
 834 | ./Data/p231/70.wav|4
 835 | ./Data/p228/42.wav|1
 836 | ./Data/p273/141.wav|19
 837 | ./Data/p256/91.wav|15
 838 | ./Data/p259/70.wav|17
 839 | ./Data/p236/69.wav|6
 840 | ./Data/p228/16.wav|1
 841 | ./Data/p270/44.wav|18
 842 | ./Data/p230/16.wav|3
 843 | ./Data/p244/97.wav|9
 844 | ./Data/p254/42.wav|14
 845 | ./Data/p225/53.wav|0
 846 | ./Data/p230/59.wav|3
 847 | ./Data/p226/140.wav|10
 848 | ./Data/p232/7.wav|12
 849 | ./Data/p229/47.wav|2
 850 | ./Data/p231/13.wav|4
 851 | ./Data/p258/49.wav|16
 852 | ./Data/p226/92.wav|10
 853 | ./Data/p227/81.wav|11
 854 | ./Data/p230/162.wav|3
 855 | ./Data/p240/20.wav|8
 856 | ./Data/p236/88.wav|6
 857 | ./Data/p236/79.wav|6
 858 | ./Data/p236/39.wav|6
 859 | ./Data/p233/97.wav|5
 860 | ./Data/p232/96.wav|12
 861 | ./Data/p273/82.wav|19
 862 | ./Data/p230/123.wav|3
 863 | ./Data/p230/126.wav|3
 864 | ./Data/p258/75.wav|16
 865 | ./Data/p232/78.wav|12
 866 | ./Data/p231/48.wav|4
 867 | ./Data/p244/110.wav|9
 868 | ./Data/p258/71.wav|16
 869 | ./Data/p256/116.wav|15
 870 | ./Data/p231/63.wav|4
 871 | ./Data/p258/26.wav|16
 872 | ./Data/p243/18.wav|13
 873 | ./Data/p243/55.wav|13
 874 | ./Data/p270/162.wav|18
 875 | ./Data/p244/33.wav|9
 876 | ./Data/p226/77.wav|10
 877 | ./Data/p270/98.wav|18
 878 | ./Data/p230/121.wav|3
 879 | ./Data/p226/94.wav|10
 880 | ./Data/p270/84.wav|18
 881 | ./Data/p270/53.wav|18
 882 | ./Data/p243/124.wav|13
 883 | ./Data/p228/86.wav|1
 884 | ./Data/p229/25.wav|2
 885 | ./Data/p230/68.wav|3
 886 | ./Data/p240/29.wav|8
 887 | ./Data/p236/63.wav|6
 888 | ./Data/p270/129.wav|18
 889 | ./Data/p229/79.wav|2
 890 | ./Data/p233/102.wav|5
 891 | ./Data/p228/34.wav|1
 892 | ./Data/p230/163.wav|3
 893 | ./Data/p228/64.wav|1
 894 | ./Data/p233/115.wav|5
 895 | ./Data/p243/88.wav|13
 896 | ./Data/p244/14.wav|9
 897 | ./Data/p243/174.wav|13
 898 | ./Data/p229/74.wav|2
 899 | ./Data/p258/27.wav|16
 900 | ./Data/p259/86.wav|17
 901 | ./Data/p273/92.wav|19
 902 | ./Data/p239/81.wav|7
 903 | ./Data/p254/109.wav|14
 904 | ./Data/p232/103.wav|12
 905 | ./Data/p230/21.wav|3
 906 | ./Data/p226/10.wav|10
 907 | ./Data/p240/2.wav|8
 908 | ./Data/p256/102.wav|15
 909 | ./Data/p240/127.wav|8
 910 | ./Data/p259/138.wav|17
 911 | ./Data/p254/123.wav|14
 912 | ./Data/p270/92.wav|18
 913 | ./Data/p254/30.wav|14
 914 | ./Data/p273/86.wav|19
 915 | ./Data/p244/106.wav|9
 916 | ./Data/p226/107.wav|10
 917 | ./Data/p240/4.wav|8
 918 | ./Data/p228/97.wav|1
 919 | ./Data/p258/32.wav|16
 920 | ./Data/p232/79.wav|12
 921 | ./Data/p259/154.wav|17
 922 | ./Data/p231/19.wav|4
 923 | ./Data/p259/91.wav|17
 924 | ./Data/p244/45.wav|9
 925 | ./Data/p240/97.wav|8
 926 | ./Data/p259/45.wav|17
 927 | ./Data/p270/197.wav|18
 928 | ./Data/p229/1.wav|2
 929 | ./Data/p259/11.wav|17
 930 | ./Data/p228/29.wav|1
 931 | ./Data/p230/72.wav|3
 932 | ./Data/p228/145.wav|1
 933 | ./Data/p244/71.wav|9
 934 | ./Data/p230/66.wav|3
 935 | ./Data/p226/51.wav|10
 936 | ./Data/p270/10.wav|18
 937 | ./Data/p254/96.wav|14
 938 | ./Data/p256/64.wav|15
 939 | ./Data/p243/65.wav|13
 940 | ./Data/p228/148.wav|1
 941 | ./Data/p243/41.wav|13
 942 | ./Data/p228/57.wav|1
 943 | ./Data/p239/92.wav|7
 944 | ./Data/p256/124.wav|15
 945 | ./Data/p259/116.wav|17
 946 | ./Data/p233/70.wav|5
 947 | ./Data/p227/1.wav|11
 948 | ./Data/p231/59.wav|4
 949 | ./Data/p243/30.wav|13
 950 | ./Data/p254/41.wav|14
 951 | ./Data/p228/123.wav|1
 952 | ./Data/p239/20.wav|7
 953 | ./Data/p229/77.wav|2
 954 | ./Data/p239/132.wav|7
 955 | ./Data/p243/144.wav|13
 956 | ./Data/p227/137.wav|11
 957 | ./Data/p239/134.wav|7
 958 | ./Data/p240/108.wav|8
 959 | ./Data/p256/118.wav|15
 960 | ./Data/p256/126.wav|15
 961 | ./Data/p226/110.wav|10
 962 | ./Data/p236/29.wav|6
 963 | ./Data/p236/74.wav|6
 964 | ./Data/p231/77.wav|4
 965 | ./Data/p256/45.wav|15
 966 | ./Data/p256/39.wav|15
 967 | ./Data/p228/66.wav|1
 968 | ./Data/p232/35.wav|12
 969 | ./Data/p273/37.wav|19
 970 | ./Data/p240/135.wav|8
 971 | ./Data/p236/73.wav|6
 972 | ./Data/p256/38.wav|15
 973 | ./Data/p243/109.wav|13
 974 | ./Data/p227/33.wav|11
 975 | ./Data/p259/87.wav|17
 976 | ./Data/p225/55.wav|0
 977 | ./Data/p243/138.wav|13
 978 | ./Data/p227/3.wav|11
 979 | ./Data/p254/74.wav|14
 980 | ./Data/p254/137.wav|14
 981 | ./Data/p228/43.wav|1
 982 | ./Data/p270/71.wav|18
 983 | ./Data/p243/56.wav|13
 984 | ./Data/p228/119.wav|1
 985 | ./Data/p244/136.wav|9
 986 | ./Data/p259/94.wav|17
 987 | ./Data/p259/120.wav|17
 988 | ./Data/p230/74.wav|3
 989 | ./Data/p227/100.wav|11
 990 | ./Data/p228/143.wav|1
 991 | ./Data/p225/98.wav|0
 992 | ./Data/p256/2.wav|15
 993 | ./Data/p273/146.wav|19
 994 | ./Data/p230/99.wav|3
 995 | ./Data/p243/20.wav|13
 996 | ./Data/p258/96.wav|16
 997 | ./Data/p226/87.wav|10
 998 | ./Data/p240/64.wav|8
 999 | ./Data/p243/114.wav|13
1000 | ./Data/p273/77.wav|19
1001 | ./Data/p256/48.wav|15
1002 | ./Data/p258/120.wav|16
1003 | ./Data/p240/111.wav|8
1004 | ./Data/p226/73.wav|10
1005 | ./Data/p229/15.wav|2
1006 | ./Data/p270/165.wav|18
1007 | ./Data/p226/124.wav|10
1008 | ./Data/p254/53.wav|14
1009 | ./Data/p239/97.wav|7
1010 | ./Data/p236/71.wav|6
1011 | ./Data/p243/66.wav|13
1012 | ./Data/p230/26.wav|3
1013 | ./Data/p233/17.wav|5
1014 | ./Data/p273/143.wav|19
1015 | ./Data/p229/6.wav|2
1016 | ./Data/p258/41.wav|16
1017 | ./Data/p240/10.wav|8
1018 | ./Data/p244/115.wav|9
1019 | ./Data/p256/8.wav|15
1020 | ./Data/p243/133.wav|13
1021 | ./Data/p236/145.wav|6
1022 | ./Data/p240/110.wav|8
1023 | ./Data/p270/100.wav|18
1024 | ./Data/p230/167.wav|3
1025 | ./Data/p270/27.wav|18
1026 | ./Data/p243/149.wav|13
1027 | ./Data/p228/139.wav|1
1028 | ./Data/p256/96.wav|15
1029 | ./Data/p230/61.wav|3
1030 | ./Data/p258/42.wav|16
1031 | ./Data/p236/94.wav|6
1032 | ./Data/p230/42.wav|3
1033 | ./Data/p270/144.wav|18
1034 | ./Data/p228/141.wav|1
1035 | ./Data/p232/4.wav|12
1036 | ./Data/p229/8.wav|2
1037 | ./Data/p230/39.wav|3
1038 | ./Data/p256/47.wav|15
1039 | ./Data/p229/54.wav|2
1040 | ./Data/p239/168.wav|7
1041 | ./Data/p227/7.wav|11
1042 | ./Data/p227/93.wav|11
1043 | ./Data/p240/13.wav|8
1044 | ./Data/p270/172.wav|18
1045 | ./Data/p243/45.wav|13
1046 | ./Data/p259/30.wav|17
1047 | ./Data/p270/116.wav|18
1048 | ./Data/p240/48.wav|8
1049 | ./Data/p227/24.wav|11
1050 | ./Data/p229/80.wav|2
1051 | ./Data/p233/2.wav|5
1052 | ./Data/p228/87.wav|1
1053 | ./Data/p240/105.wav|8
1054 | ./Data/p239/60.wav|7
1055 | ./Data/p244/39.wav|9
1056 | ./Data/p240/124.wav|8
1057 | ./Data/p259/145.wav|17
1058 | ./Data/p227/76.wav|11
1059 | ./Data/p254/58.wav|14
1060 | ./Data/p230/156.wav|3
1061 | ./Data/p229/42.wav|2
1062 | ./Data/p273/68.wav|19
1063 | ./Data/p228/146.wav|1
1064 | ./Data/p236/165.wav|6
1065 | ./Data/p229/34.wav|2
1066 | ./Data/p239/123.wav|7
1067 | ./Data/p273/121.wav|19
1068 | ./Data/p270/176.wav|18
1069 | ./Data/p258/74.wav|16
1070 | ./Data/p254/84.wav|14
1071 | ./Data/p259/157.wav|17
1072 | ./Data/p258/130.wav|16
1073 | ./Data/p244/18.wav|9
1074 | ./Data/p229/59.wav|2
1075 | ./Data/p229/10.wav|2
1076 | ./Data/p273/89.wav|19
1077 | ./Data/p259/23.wav|17
1078 | ./Data/p256/6.wav|15
1079 | ./Data/p227/8.wav|11
1080 | ./Data/p258/59.wav|16
1081 | ./Data/p232/91.wav|12
1082 | ./Data/p258/137.wav|16
1083 | ./Data/p258/122.wav|16
1084 | ./Data/p230/89.wav|3
1085 | ./Data/p232/58.wav|12
1086 | ./Data/p231/11.wav|4
1087 | ./Data/p273/120.wav|19
1088 | ./Data/p232/39.wav|12
1089 | ./Data/p236/44.wav|6
1090 | ./Data/p254/12.wav|14
1091 | ./Data/p270/95.wav|18
1092 | ./Data/p270/153.wav|18
1093 | ./Data/p230/164.wav|3
1094 | ./Data/p225/30.wav|0
1095 | ./Data/p240/126.wav|8
1096 | ./Data/p230/54.wav|3
1097 | ./Data/p270/87.wav|18
1098 | ./Data/p225/14.wav|0
1099 | ./Data/p231/145.wav|4
1100 | ./Data/p254/81.wav|14
1101 | ./Data/p244/55.wav|9
1102 | ./Data/p259/3.wav|17
1103 | ./Data/p273/50.wav|19
1104 | ./Data/p228/84.wav|1
1105 | ./Data/p244/3.wav|9
1106 | ./Data/p239/55.wav|7
1107 | ./Data/p232/5.wav|12
1108 | ./Data/p229/111.wav|2
1109 | ./Data/p236/141.wav|6
1110 | ./Data/p233/54.wav|5
1111 | ./Data/p240/88.wav|8
1112 | ./Data/p236/16.wav|6
1113 | ./Data/p239/154.wav|7
1114 | ./Data/p240/72.wav|8
1115 | ./Data/p236/75.wav|6
1116 | ./Data/p230/166.wav|3
1117 | ./Data/p231/122.wav|4
1118 | ./Data/p273/24.wav|19
1119 | ./Data/p233/30.wav|5
1120 | ./Data/p226/9.wav|10
1121 | ./Data/p240/65.wav|8
1122 | ./Data/p228/80.wav|1
1123 | ./Data/p232/46.wav|12
1124 | ./Data/p239/109.wav|7
1125 | ./Data/p231/67.wav|4
1126 | ./Data/p233/67.wav|5
1127 | ./Data/p228/162.wav|1
1128 | ./Data/p229/134.wav|2
1129 | ./Data/p239/27.wav|7
1130 | ./Data/p227/145.wav|11
1131 | ./Data/p225/67.wav|0
1132 | ./Data/p232/99.wav|12
1133 | ./Data/p270/140.wav|18
1134 | ./Data/p225/70.wav|0
1135 | ./Data/p259/21.wav|17
1136 | ./Data/p230/28.wav|3
1137 | ./Data/p230/80.wav|3
1138 | ./Data/p243/34.wav|13
1139 | ./Data/p254/61.wav|14
1140 | ./Data/p236/58.wav|6
1141 | ./Data/p239/21.wav|7
1142 | ./Data/p230/91.wav|3
1143 | ./Data/p256/68.wav|15
1144 | ./Data/p225/21.wav|0
1145 | ./Data/p233/49.wav|5
1146 | ./Data/p236/114.wav|6
1147 | ./Data/p228/134.wav|1
1148 | ./Data/p231/114.wav|4
1149 | ./Data/p239/18.wav|7
1150 | ./Data/p227/132.wav|11
1151 | ./Data/p236/115.wav|6
1152 | ./Data/p254/99.wav|14
1153 | ./Data/p243/143.wav|13
1154 | ./Data/p270/49.wav|18
1155 | ./Data/p239/152.wav|7
1156 | ./Data/p232/120.wav|12
1157 | ./Data/p256/25.wav|15
1158 | ./Data/p229/116.wav|2
1159 | ./Data/p239/130.wav|7
1160 | ./Data/p254/124.wav|14
1161 | ./Data/p270/118.wav|18
1162 | ./Data/p244/46.wav|9
1163 | ./Data/p231/105.wav|4
1164 | ./Data/p231/115.wav|4
1165 | ./Data/p239/144.wav|7
1166 | ./Data/p226/39.wav|10
1167 | ./Data/p233/78.wav|5
1168 | ./Data/p227/53.wav|11
1169 | ./Data/p239/146.wav|7
1170 | ./Data/p256/77.wav|15
1171 | ./Data/p259/37.wav|17
1172 | ./Data/p258/36.wav|16
1173 | ./Data/p254/13.wav|14
1174 | ./Data/p229/69.wav|2
1175 | ./Data/p231/90.wav|4
1176 | ./Data/p226/84.wav|10
1177 | ./Data/p259/48.wav|17
1178 | ./Data/p233/88.wav|5
1179 | ./Data/p228/153.wav|1
1180 | ./Data/p254/43.wav|14
1181 | ./Data/p231/97.wav|4
1182 | ./Data/p273/44.wav|19
1183 | ./Data/p233/27.wav|5
1184 | ./Data/p232/90.wav|12
1185 | ./Data/p254/36.wav|14
1186 | ./Data/p232/27.wav|12
1187 | ./Data/p230/113.wav|3
1188 | ./Data/p254/130.wav|14
1189 | ./Data/p254/62.wav|14
1190 | ./Data/p239/118.wav|7
1191 | ./Data/p230/109.wav|3
1192 | ./Data/p227/102.wav|11
1193 | ./Data/p226/48.wav|10
1194 | ./Data/p230/175.wav|3
1195 | ./Data/p231/60.wav|4
1196 | ./Data/p259/105.wav|17
1197 | ./Data/p233/28.wav|5
1198 | ./Data/p229/36.wav|2
1199 | ./Data/p256/111.wav|15
1200 | ./Data/p230/133.wav|3
1201 | ./Data/p233/125.wav|5
1202 | ./Data/p228/59.wav|1
1203 | ./Data/p239/58.wav|7
1204 | ./Data/p273/116.wav|19
1205 | ./Data/p230/97.wav|3
1206 | ./Data/p273/88.wav|19
1207 | ./Data/p228/93.wav|1
1208 | ./Data/p259/81.wav|17
1209 | ./Data/p228/144.wav|1
1210 | ./Data/p230/32.wav|3
1211 | ./Data/p240/6.wav|8
1212 | ./Data/p230/17.wav|3
1213 | ./Data/p259/98.wav|17
1214 | ./Data/p227/75.wav|11
1215 | ./Data/p231/26.wav|4
1216 | ./Data/p231/103.wav|4
1217 | ./Data/p236/67.wav|6
1218 | ./Data/p270/107.wav|18
1219 | ./Data/p226/24.wav|10
1220 | ./Data/p273/34.wav|19
1221 | ./Data/p236/90.wav|6
1222 | ./Data/p256/14.wav|15
1223 | ./Data/p236/140.wav|6
1224 | ./Data/p273/39.wav|19
1225 | ./Data/p270/163.wav|18
1226 | ./Data/p239/77.wav|7
1227 | ./Data/p230/148.wav|3
1228 | ./Data/p273/113.wav|19
1229 | ./Data/p254/140.wav|14
1230 | ./Data/p239/46.wav|7
1231 | ./Data/p243/51.wav|13
1232 | ./Data/p231/10.wav|4
1233 | ./Data/p231/104.wav|4
1234 | ./Data/p270/132.wav|18
1235 | ./Data/p228/108.wav|1
1236 | ./Data/p233/39.wav|5
1237 | ./Data/p259/130.wav|17
1238 | ./Data/p239/85.wav|7
1239 | ./Data/p240/37.wav|8
1240 | ./Data/p270/58.wav|18
1241 | ./Data/p243/78.wav|13
1242 | ./Data/p273/61.wav|19
1243 | ./Data/p230/144.wav|3
1244 | ./Data/p233/21.wav|5
1245 | ./Data/p225/35.wav|0
1246 | ./Data/p228/158.wav|1
1247 | ./Data/p259/26.wav|17
1248 | ./Data/p230/33.wav|3
1249 | ./Data/p258/128.wav|16
1250 | ./Data/p233/61.wav|5
1251 | ./Data/p225/97.wav|0
1252 | ./Data/p259/143.wav|17
1253 | ./Data/p226/50.wav|10
1254 | ./Data/p243/71.wav|13
1255 | ./Data/p230/22.wav|3
1256 | ./Data/p226/58.wav|10
1257 | ./Data/p239/110.wav|7
1258 | ./Data/p258/136.wav|16
1259 | ./Data/p226/102.wav|10
1260 | ./Data/p258/88.wav|16
1261 | ./Data/p233/94.wav|5
1262 | ./Data/p258/77.wav|16
1263 | ./Data/p231/2.wav|4
1264 | ./Data/p273/40.wav|19
1265 | ./Data/p239/133.wav|7
1266 | ./Data/p270/33.wav|18
1267 | ./Data/p254/132.wav|14
1268 | ./Data/p270/99.wav|18
1269 | ./Data/p227/84.wav|11
1270 | ./Data/p226/132.wav|10
1271 | ./Data/p239/165.wav|7
1272 | ./Data/p270/23.wav|18
1273 | ./Data/p270/41.wav|18
1274 | ./Data/p236/28.wav|6
1275 | ./Data/p231/76.wav|4
1276 | ./Data/p231/28.wav|4
1277 | ./Data/p236/56.wav|6
1278 | ./Data/p236/146.wav|6
1279 | ./Data/p244/125.wav|9
1280 | ./Data/p256/55.wav|15
1281 | ./Data/p232/40.wav|12
1282 | ./Data/p239/64.wav|7
1283 | ./Data/p240/130.wav|8
1284 | ./Data/p239/41.wav|7
1285 | ./Data/p240/138.wav|8
1286 | ./Data/p226/118.wav|10
1287 | ./Data/p228/62.wav|1
1288 | ./Data/p244/16.wav|9
1289 | ./Data/p244/20.wav|9
1290 | ./Data/p226/125.wav|10
1291 | ./Data/p270/74.wav|18
1292 | ./Data/p229/129.wav|2
1293 | ./Data/p227/142.wav|11
1294 | ./Data/p228/38.wav|1
1295 | ./Data/p258/97.wav|16
1296 | ./Data/p233/77.wav|5
1297 | ./Data/p232/84.wav|12
1298 | ./Data/p229/17.wav|2
1299 | ./Data/p227/18.wav|11
1300 | ./Data/p239/94.wav|7
1301 | ./Data/p239/1.wav|7
1302 | ./Data/p225/52.wav|0
1303 | ./Data/p270/82.wav|18
1304 | ./Data/p232/53.wav|12
1305 | ./Data/p258/51.wav|16
1306 | ./Data/p258/132.wav|16
1307 | ./Data/p229/66.wav|2
1308 | ./Data/p270/19.wav|18
1309 | ./Data/p227/88.wav|11
1310 | ./Data/p231/96.wav|4
1311 | ./Data/p239/72.wav|7
1312 | ./Data/p225/73.wav|0
1313 | ./Data/p240/146.wav|8
1314 | ./Data/p236/97.wav|6
1315 | ./Data/p227/43.wav|11
1316 | ./Data/p232/119.wav|12
1317 | ./Data/p231/53.wav|4
1318 | ./Data/p239/42.wav|7
1319 | ./Data/p259/115.wav|17
1320 | ./Data/p244/105.wav|9
1321 | ./Data/p256/33.wav|15
1322 | ./Data/p231/100.wav|4
1323 | ./Data/p240/8.wav|8
1324 | ./Data/p256/57.wav|15
1325 | ./Data/p227/130.wav|11
1326 | ./Data/p226/30.wav|10
1327 | ./Data/p233/80.wav|5
1328 | ./Data/p232/17.wav|12
1329 | ./Data/p259/167.wav|17
1330 | ./Data/p227/122.wav|11
1331 | ./Data/p239/128.wav|7
1332 | ./Data/p231/133.wav|4
1333 | ./Data/p273/129.wav|19
1334 | ./Data/p243/15.wav|13
1335 | ./Data/p243/44.wav|13
1336 | ./Data/p259/161.wav|17
1337 | ./Data/p243/94.wav|13
1338 | ./Data/p244/62.wav|9
1339 | ./Data/p270/180.wav|18
1340 | ./Data/p258/126.wav|16
1341 | ./Data/p229/137.wav|2
1342 | ./Data/p233/105.wav|5
1343 | ./Data/p244/79.wav|9
1344 | ./Data/p254/46.wav|14
1345 | ./Data/p240/95.wav|8
1346 | ./Data/p259/135.wav|17
1347 | ./Data/p259/52.wav|17
1348 | ./Data/p229/68.wav|2
1349 | ./Data/p254/33.wav|14
1350 | ./Data/p230/83.wav|3
1351 | ./Data/p256/89.wav|15
1352 | ./Data/p254/90.wav|14
1353 | ./Data/p270/182.wav|18
1354 | ./Data/p226/18.wav|10
1355 | ./Data/p270/145.wav|18
1356 | ./Data/p231/128.wav|4
1357 | ./Data/p239/140.wav|7
1358 | ./Data/p228/100.wav|1
1359 | ./Data/p227/49.wav|11
1360 | ./Data/p240/53.wav|8
1361 | ./Data/p258/108.wav|16
1362 | ./Data/p226/83.wav|10
1363 | ./Data/p270/106.wav|18
1364 | ./Data/p243/11.wav|13
1365 | ./Data/p229/12.wav|2
1366 | ./Data/p228/7.wav|1
1367 | ./Data/p243/8.wav|13
1368 | ./Data/p227/128.wav|11
1369 | ./Data/p230/118.wav|3
1370 | ./Data/p227/78.wav|11
1371 | ./Data/p244/30.wav|9
1372 | ./Data/p231/98.wav|4
1373 | ./Data/p230/38.wav|3
1374 | ./Data/p244/47.wav|9
1375 | ./Data/p270/138.wav|18
1376 | ./Data/p259/109.wav|17
1377 | ./Data/p270/112.wav|18
1378 | ./Data/p227/82.wav|11
1379 | ./Data/p228/161.wav|1
1380 | ./Data/p273/127.wav|19
1381 | ./Data/p232/72.wav|12
1382 | ./Data/p227/95.wav|11
1383 | ./Data/p236/105.wav|6
1384 | ./Data/p239/52.wav|7
1385 | ./Data/p273/135.wav|19
1386 | ./Data/p236/136.wav|6
1387 | ./Data/p228/113.wav|1
1388 | ./Data/p229/56.wav|2
1389 | ./Data/p240/34.wav|8
1390 | ./Data/p230/79.wav|3
1391 | ./Data/p232/48.wav|12
1392 | ./Data/p240/101.wav|8
1393 | ./Data/p229/112.wav|2
1394 | ./Data/p273/46.wav|19
1395 | ./Data/p273/27.wav|19
1396 | ./Data/p239/103.wav|7
1397 | ./Data/p259/117.wav|17
1398 | ./Data/p230/37.wav|3
1399 | ./Data/p233/138.wav|5
1400 | ./Data/p228/125.wav|1
1401 | ./Data/p230/115.wav|3
1402 | ./Data/p240/42.wav|8
1403 | ./Data/p231/99.wav|4
1404 | ./Data/p236/54.wav|6
1405 | ./Data/p233/104.wav|5
1406 | ./Data/p270/4.wav|18
1407 | ./Data/p226/122.wav|10
1408 | ./Data/p230/56.wav|3
1409 | ./Data/p244/58.wav|9
1410 | ./Data/p229/133.wav|2
1411 | ./Data/p270/64.wav|18
1412 | ./Data/p225/88.wav|0
1413 | ./Data/p240/104.wav|8
1414 | ./Data/p244/78.wav|9
1415 | ./Data/p254/113.wav|14
1416 | ./Data/p259/144.wav|17
1417 | ./Data/p236/100.wav|6
1418 | ./Data/p230/81.wav|3
1419 | ./Data/p259/53.wav|17
1420 | ./Data/p239/155.wav|7
1421 | ./Data/p236/148.wav|6
1422 | ./Data/p270/8.wav|18
1423 | ./Data/p225/90.wav|0
1424 | ./Data/p236/64.wav|6
1425 | ./Data/p236/159.wav|6
1426 | ./Data/p232/63.wav|12
1427 | ./Data/p244/2.wav|9
1428 | ./Data/p258/28.wav|16
1429 | ./Data/p259/5.wav|17
1430 | ./Data/p225/42.wav|0
1431 | ./Data/p256/49.wav|15
1432 | ./Data/p233/24.wav|5
1433 | ./Data/p270/146.wav|18
1434 | ./Data/p243/131.wav|13
1435 | ./Data/p229/91.wav|2
1436 | ./Data/p229/76.wav|2
1437 | ./Data/p227/22.wav|11
1438 | ./Data/p244/59.wav|9
1439 | ./Data/p236/17.wav|6
1440 | ./Data/p240/32.wav|8
1441 | ./Data/p232/23.wav|12
1442 | ./Data/p230/20.wav|3
1443 | ./Data/p232/111.wav|12
1444 | ./Data/p230/159.wav|3
1445 | ./Data/p244/15.wav|9
1446 | ./Data/p229/86.wav|2
1447 | ./Data/p240/54.wav|8
1448 | ./Data/p229/132.wav|2
1449 | ./Data/p239/126.wav|7
1450 | ./Data/p240/91.wav|8
1451 | ./Data/p244/51.wav|9
1452 | ./Data/p254/19.wav|14
1453 | ./Data/p244/32.wav|9
1454 | ./Data/p258/114.wav|16
1455 | ./Data/p254/106.wav|14
1456 | ./Data/p243/111.wav|13
1457 | ./Data/p226/106.wav|10
1458 | ./Data/p244/26.wav|9
1459 | ./Data/p225/57.wav|0
1460 | ./Data/p243/24.wav|13
1461 | ./Data/p259/127.wav|17
1462 | ./Data/p256/50.wav|15
1463 | ./Data/p239/100.wav|7
1464 | ./Data/p273/10.wav|19
1465 | ./Data/p229/2.wav|2
1466 | ./Data/p270/70.wav|18
1467 | ./Data/p254/95.wav|14
1468 | ./Data/p256/120.wav|15
1469 | ./Data/p233/107.wav|5
1470 | ./Data/p226/90.wav|10
1471 | ./Data/p258/55.wav|16
1472 | ./Data/p233/99.wav|5
1473 | ./Data/p230/6.wav|3
1474 | ./Data/p273/131.wav|19
1475 | ./Data/p273/52.wav|19
1476 | ./Data/p236/158.wav|6
1477 | ./Data/p232/62.wav|12
1478 | ./Data/p233/20.wav|5
1479 | ./Data/p270/90.wav|18
1480 | ./Data/p240/11.wav|8
1481 | ./Data/p258/66.wav|16
1482 | ./Data/p258/65.wav|16
1483 | ./Data/p270/94.wav|18
1484 | ./Data/p270/9.wav|18
1485 | ./Data/p228/82.wav|1
1486 | ./Data/p236/96.wav|6
1487 | ./Data/p229/33.wav|2
1488 | ./Data/p229/19.wav|2
1489 | ./Data/p239/54.wav|7
1490 | ./Data/p232/106.wav|12
1491 | ./Data/p231/138.wav|4
1492 | ./Data/p230/57.wav|3
1493 | ./Data/p270/89.wav|18
1494 | ./Data/p273/95.wav|19
1495 | ./Data/p231/131.wav|4
1496 | ./Data/p236/107.wav|6
1497 | ./Data/p228/122.wav|1
1498 | ./Data/p226/109.wav|10
1499 | ./Data/p270/117.wav|18
1500 | ./Data/p230/110.wav|3
1501 | ./Data/p270/37.wav|18
1502 | ./Data/p225/29.wav|0
1503 | ./Data/p233/8.wav|5
1504 | ./Data/p227/4.wav|11
1505 | ./Data/p232/97.wav|12
1506 | ./Data/p243/14.wav|13
1507 | ./Data/p254/91.wav|14
1508 | ./Data/p256/62.wav|15
1509 | ./Data/p229/110.wav|2
1510 | ./Data/p233/34.wav|5
1511 | ./Data/p226/81.wav|10
1512 | ./Data/p230/29.wav|3
1513 | ./Data/p240/84.wav|8
1514 | ./Data/p270/201.wav|18
1515 | ./Data/p239/157.wav|7
1516 | ./Data/p270/158.wav|18
1517 | ./Data/p236/80.wav|6
1518 | ./Data/p232/54.wav|12
1519 | ./Data/p239/29.wav|7
1520 | ./Data/p225/33.wav|0
1521 | ./Data/p273/7.wav|19
1522 | ./Data/p273/98.wav|19
1523 | ./Data/p227/63.wav|11
1524 | ./Data/p230/174.wav|3
1525 | ./Data/p270/28.wav|18
1526 | ./Data/p233/13.wav|5
1527 | ./Data/p273/99.wav|19
1528 | ./Data/p229/81.wav|2
1529 | ./Data/p273/124.wav|19
1530 | ./Data/p230/129.wav|3
1531 | ./Data/p259/133.wav|17
1532 | ./Data/p270/24.wav|18
1533 | ./Data/p226/35.wav|10
1534 | ./Data/p236/118.wav|6
1535 | ./Data/p254/121.wav|14
1536 | ./Data/p270/120.wav|18
1537 | ./Data/p231/30.wav|4
1538 | ./Data/p240/102.wav|8
1539 | ./Data/p243/53.wav|13
1540 | ./Data/p230/47.wav|3
1541 | ./Data/p233/55.wav|5
1542 | ./Data/p226/11.wav|10
1543 | ./Data/p239/120.wav|7
1544 | ./Data/p226/49.wav|10
1545 | ./Data/p239/44.wav|7
1546 | ./Data/p244/140.wav|9
1547 | ./Data/p258/63.wav|16
1548 | ./Data/p232/52.wav|12
1549 | ./Data/p273/109.wav|19
1550 | ./Data/p259/72.wav|17
1551 | ./Data/p259/164.wav|17
1552 | ./Data/p256/78.wav|15
1553 | ./Data/p243/107.wav|13
1554 | ./Data/p258/62.wav|16
1555 | ./Data/p239/31.wav|7
1556 | ./Data/p256/41.wav|15
1557 | ./Data/p273/63.wav|19
1558 | ./Data/p258/112.wav|16
1559 | ./Data/p243/116.wav|13
1560 | ./Data/p254/29.wav|14
1561 | ./Data/p229/45.wav|2
1562 | ./Data/p244/101.wav|9
1563 | ./Data/p232/34.wav|12
1564 | ./Data/p243/154.wav|13
1565 | ./Data/p231/33.wav|4
1566 | ./Data/p243/35.wav|13
1567 | ./Data/p236/38.wav|6
1568 | ./Data/p270/16.wav|18
1569 | ./Data/p270/187.wav|18
1570 | ./Data/p239/114.wav|7
1571 | ./Data/p244/24.wav|9
1572 | ./Data/p228/75.wav|1
1573 | ./Data/p226/26.wav|10
1574 | ./Data/p259/136.wav|17
1575 | ./Data/p236/147.wav|6
1576 | ./Data/p239/135.wav|7
1577 | ./Data/p270/43.wav|18
1578 | ./Data/p244/132.wav|9
1579 | ./Data/p243/129.wav|13
1580 | ./Data/p236/9.wav|6
1581 | ./Data/p232/109.wav|12
1582 | ./Data/p225/84.wav|0
1583 | ./Data/p227/27.wav|11
1584 | ./Data/p259/8.wav|17
1585 | ./Data/p259/67.wav|17
1586 | ./Data/p239/57.wav|7
1587 | ./Data/p243/69.wav|13
1588 | ./Data/p231/62.wav|4
1589 | ./Data/p259/140.wav|17
1590 | ./Data/p227/66.wav|11
1591 | ./Data/p230/44.wav|3
1592 | ./Data/p229/63.wav|2
1593 | ./Data/p256/4.wav|15
1594 | ./Data/p258/24.wav|16
1595 | ./Data/p240/80.wav|8
1596 | ./Data/p270/72.wav|18
1597 | ./Data/p240/47.wav|8
1598 | ./Data/p229/98.wav|2
1599 | ./Data/p244/111.wav|9
1600 | ./Data/p231/111.wav|4
1601 | ./Data/p243/91.wav|13
1602 | ./Data/p239/36.wav|7
1603 | ./Data/p259/103.wav|17
1604 | ./Data/p232/2.wav|12
1605 | ./Data/p236/3.wav|6
1606 | ./Data/p236/57.wav|6
1607 | ./Data/p233/109.wav|5
1608 | ./Data/p236/122.wav|6
1609 | ./Data/p270/76.wav|18
1610 | ./Data/p243/167.wav|13
1611 | ./Data/p228/20.wav|1
1612 | ./Data/p243/72.wav|13
1613 | ./Data/p239/2.wav|7
1614 | ./Data/p226/21.wav|10
1615 | ./Data/p256/43.wav|15
1616 | ./Data/p259/129.wav|17
1617 | ./Data/p231/15.wav|4
1618 | ./Data/p231/85.wav|4
1619 | ./Data/p226/29.wav|10
1620 | ./Data/p230/131.wav|3
1621 | ./Data/p259/97.wav|17
1622 | ./Data/p240/68.wav|8
1623 | ./Data/p233/84.wav|5
1624 | ./Data/p236/10.wav|6
1625 | ./Data/p244/120.wav|9
1626 | ./Data/p270/18.wav|18
1627 | ./Data/p231/24.wav|4
1628 | ./Data/p256/37.wav|15
1629 | ./Data/p233/11.wav|5
1630 | ./Data/p230/93.wav|3
1631 | ./Data/p230/73.wav|3
1632 | ./Data/p239/66.wav|7
1633 | ./Data/p230/40.wav|3
1634 | ./Data/p228/13.wav|1
1635 | ./Data/p231/49.wav|4
1636 | ./Data/p270/62.wav|18
1637 | ./Data/p236/78.wav|6
1638 | ./Data/p258/73.wav|16
1639 | ./Data/p236/35.wav|6
1640 | ./Data/p254/120.wav|14
1641 | ./Data/p258/53.wav|16
1642 | ./Data/p227/16.wav|11
1643 | ./Data/p232/33.wav|12
1644 | ./Data/p256/42.wav|15
1645 | ./Data/p233/68.wav|5
1646 | ./Data/p225/74.wav|0
1647 | ./Data/p244/127.wav|9
1648 | ./Data/p243/118.wav|13
1649 | ./Data/p273/128.wav|19
1650 | ./Data/p239/7.wav|7
1651 | ./Data/p243/50.wav|13
1652 | ./Data/p226/23.wav|10
1653 | ./Data/p270/199.wav|18
1654 | ./Data/p254/45.wav|14
1655 | ./Data/p254/11.wav|14
1656 | ./Data/p244/66.wav|9
1657 | ./Data/p270/152.wav|18
1658 | ./Data/p227/131.wav|11
1659 | ./Data/p270/38.wav|18
1660 | ./Data/p229/57.wav|2
1661 | ./Data/p227/35.wav|11
1662 | ./Data/p244/7.wav|9
1663 | ./Data/p226/32.wav|10
1664 | ./Data/p230/152.wav|3
1665 | ./Data/p239/161.wav|7
1666 | ./Data/p256/123.wav|15
1667 | ./Data/p231/14.wav|4
1668 | ./Data/p243/38.wav|13
1669 | ./Data/p229/102.wav|2
1670 | ./Data/p229/38.wav|2
1671 | ./Data/p233/116.wav|5
1672 | ./Data/p254/35.wav|14
1673 | ./Data/p254/118.wav|14
1674 | ./Data/p225/15.wav|0
1675 | ./Data/p230/132.wav|3
1676 | ./Data/p273/84.wav|19
1677 | ./Data/p254/122.wav|14
1678 | ./Data/p273/3.wav|19
1679 | ./Data/p270/68.wav|18
1680 | ./Data/p232/42.wav|12
1681 | ./Data/p225/93.wav|0
1682 | ./Data/p227/34.wav|11
1683 | ./Data/p270/22.wav|18
1684 | ./Data/p231/4.wav|4
1685 | ./Data/p227/125.wav|11
1686 | ./Data/p244/95.wav|9
1687 | ./Data/p236/18.wav|6
1688 | ./Data/p273/25.wav|19
1689 | ./Data/p259/169.wav|17
1690 | ./Data/p233/56.wav|5
1691 | ./Data/p270/203.wav|18
1692 | ./Data/p259/41.wav|17
1693 | ./Data/p233/38.wav|5
1694 | ./Data/p229/22.wav|2
1695 | ./Data/p256/17.wav|15
1696 | ./Data/p270/3.wav|18
1697 | ./Data/p231/5.wav|4
1698 | ./Data/p240/60.wav|8
1699 | ./Data/p227/21.wav|11
1700 | ./Data/p259/1.wav|17
1701 | ./Data/p259/4.wav|17
1702 | ./Data/p232/11.wav|12
1703 | ./Data/p259/114.wav|17
1704 | ./Data/p226/45.wav|10
1705 | ./Data/p236/27.wav|6
1706 | ./Data/p239/47.wav|7
1707 | ./Data/p244/85.wav|9
1708 | ./Data/p243/87.wav|13
1709 | ./Data/p258/89.wav|16
1710 | ./Data/p233/57.wav|5
1711 | ./Data/p228/78.wav|1
1712 | ./Data/p256/60.wav|15
1713 | ./Data/p232/83.wav|12
1714 | ./Data/p232/88.wav|12
1715 | ./Data/p231/120.wav|4
1716 | ./Data/p226/101.wav|10
1717 | ./Data/p236/102.wav|6
1718 | ./Data/p226/123.wav|10
1719 | ./Data/p259/85.wav|17
1720 | ./Data/p227/124.wav|11
1721 | ./Data/p259/80.wav|17
1722 | ./Data/p227/10.wav|11
1723 | ./Data/p233/26.wav|5
1724 | ./Data/p273/75.wav|19
1725 | ./Data/p243/73.wav|13
1726 | ./Data/p244/22.wav|9
1727 | ./Data/p243/126.wav|13
1728 | ./Data/p244/108.wav|9
1729 | ./Data/p243/134.wav|13
1730 | ./Data/p226/100.wav|10
1731 | ./Data/p231/123.wav|4
1732 | ./Data/p228/47.wav|1
1733 | ./Data/p243/42.wav|13
1734 | ./Data/p233/131.wav|5
1735 | ./Data/p273/2.wav|19
1736 | ./Data/p254/24.wav|14
1737 | ./Data/p236/123.wav|6
1738 | ./Data/p240/24.wav|8
1739 | ./Data/p244/63.wav|9
1740 | ./Data/p236/149.wav|6
1741 | ./Data/p236/83.wav|6
1742 | ./Data/p258/131.wav|16
1743 | ./Data/p243/120.wav|13
1744 | ./Data/p259/159.wav|17
1745 | ./Data/p258/8.wav|16
1746 | ./Data/p258/34.wav|16
1747 | ./Data/p243/33.wav|13
1748 | ./Data/p256/18.wav|15
1749 | ./Data/p232/73.wav|12
1750 | ./Data/p244/49.wav|9
1751 | ./Data/p258/12.wav|16
1752 | ./Data/p225/18.wav|0
1753 | ./Data/p258/68.wav|16
1754 | ./Data/p270/134.wav|18
1755 | ./Data/p228/54.wav|1
1756 | ./Data/p236/139.wav|6
1757 | ./Data/p225/6.wav|0
1758 | ./Data/p259/57.wav|17
1759 | ./Data/p243/70.wav|13
1760 | ./Data/p240/122.wav|8
1761 | ./Data/p259/69.wav|17
1762 | ./Data/p258/124.wav|16
1763 | ./Data/p226/138.wav|10
1764 | ./Data/p231/51.wav|4
1765 | ./Data/p259/126.wav|17
1766 | ./Data/p227/119.wav|11
1767 | ./Data/p254/136.wav|14
1768 | ./Data/p240/107.wav|8
1769 | ./Data/p254/4.wav|14
1770 | ./Data/p228/117.wav|1
1771 | ./Data/p244/92.wav|9
1772 | ./Data/p239/151.wav|7
1773 | ./Data/p259/131.wav|17
1774 | ./Data/p273/96.wav|19
1775 | ./Data/p254/69.wav|14
1776 | ./Data/p259/16.wav|17
1777 | ./Data/p244/86.wav|9
1778 | ./Data/p236/30.wav|6
1779 | ./Data/p230/34.wav|3
1780 | ./Data/p230/142.wav|3
1781 | ./Data/p244/37.wav|9
1782 | ./Data/p239/40.wav|7
1783 | ./Data/p232/87.wav|12
1784 | ./Data/p270/115.wav|18
1785 | ./Data/p232/124.wav|12
1786 | ./Data/p233/127.wav|5
1787 | ./Data/p228/70.wav|1
1788 | ./Data/p254/66.wav|14
1789 | ./Data/p232/16.wav|12
1790 | ./Data/p256/109.wav|15
1791 | ./Data/p243/169.wav|13
1792 | ./Data/p228/112.wav|1
1793 | ./Data/p254/82.wav|14
1794 | ./Data/p231/119.wav|4
1795 | ./Data/p236/59.wav|6
1796 | ./Data/p239/69.wav|7
1797 | ./Data/p225/12.wav|0
1798 | ./Data/p232/18.wav|12
1799 | ./Data/p229/32.wav|2
1800 | ./Data/p228/126.wav|1
1801 | ./Data/p270/171.wav|18
1802 | ./Data/p236/13.wav|6
1803 | ./Data/p228/12.wav|1
1804 | ./Data/p228/96.wav|1
1805 | ./Data/p256/11.wav|15
1806 | ./Data/p233/83.wav|5
1807 | ./Data/p256/99.wav|15
1808 | ./Data/p225/69.wav|0
1809 | ./Data/p254/7.wav|14
1810 | ./Data/p227/59.wav|11
1811 | ./Data/p273/136.wav|19
1812 | ./Data/p239/3.wav|7
1813 | ./Data/p256/119.wav|15
1814 | ./Data/p226/99.wav|10
1815 | ./Data/p256/56.wav|15
1816 | ./Data/p243/82.wav|13
1817 | ./Data/p227/69.wav|11
1818 | ./Data/p273/29.wav|19
1819 | ./Data/p233/100.wav|5
1820 | ./Data/p230/173.wav|3
1821 | ./Data/p240/132.wav|8
1822 | ./Data/p239/143.wav|7
1823 | ./Data/p231/40.wav|4
1824 | ./Data/p256/10.wav|15
1825 | ./Data/p229/75.wav|2
1826 | ./Data/p240/15.wav|8
1827 | ./Data/p228/102.wav|1
1828 | ./Data/p270/52.wav|18
1829 | ./Data/p270/7.wav|18
1830 | ./Data/p270/164.wav|18
1831 | ./Data/p233/91.wav|5
1832 | ./Data/p244/27.wav|9
1833 | ./Data/p244/48.wav|9
1834 | ./Data/p239/24.wav|7
1835 | ./Data/p226/113.wav|10
1836 | ./Data/p227/72.wav|11
1837 | ./Data/p270/67.wav|18
1838 | ./Data/p231/25.wav|4
1839 | ./Data/p229/120.wav|2
1840 | ./Data/p273/67.wav|19
1841 | ./Data/p230/67.wav|3
1842 | ./Data/p227/120.wav|11
1843 | ./Data/p239/121.wav|7
1844 | ./Data/p228/88.wav|1
1845 | ./Data/p254/15.wav|14
1846 | ./Data/p270/114.wav|18
1847 | ./Data/p254/14.wav|14
1848 | ./Data/p259/75.wav|17
1849 | ./Data/p236/126.wav|6
1850 | ./Data/p228/92.wav|1
1851 | ./Data/p230/127.wav|3
1852 | ./Data/p229/93.wav|2
1853 | ./Data/p233/82.wav|5
1854 | ./Data/p239/122.wav|7
1855 | ./Data/p229/72.wav|2
1856 | ./Data/p232/131.wav|12
1857 | ./Data/p239/67.wav|7
1858 | ./Data/p225/36.wav|0
1859 | ./Data/p254/3.wav|14
1860 | ./Data/p244/109.wav|9
1861 | ./Data/p230/112.wav|3
1862 | ./Data/p230/5.wav|3
1863 | ./Data/p256/87.wav|15
1864 | ./Data/p232/15.wav|12
1865 | ./Data/p244/67.wav|9
1866 | ./Data/p236/48.wav|6
1867 | ./Data/p232/110.wav|12
1868 | ./Data/p243/156.wav|13
1869 | ./Data/p231/140.wav|4
1870 | ./Data/p239/89.wav|7
1871 | ./Data/p229/53.wav|2
1872 | ./Data/p256/97.wav|15
1873 | ./Data/p256/79.wav|15
1874 | ./Data/p236/6.wav|6
1875 | ./Data/p236/106.wav|6
1876 | ./Data/p227/15.wav|11
1877 | ./Data/p273/20.wav|19
1878 | ./Data/p239/49.wav|7
1879 | ./Data/p254/134.wav|14
1880 | ./Data/p228/4.wav|1
1881 | ./Data/p227/117.wav|11
1882 | ./Data/p259/7.wav|17
1883 | ./Data/p258/91.wav|16
1884 | ./Data/p259/128.wav|17
1885 | ./Data/p236/61.wav|6
1886 | ./Data/p230/165.wav|3
1887 | ./Data/p225/20.wav|0
1888 | ./Data/p232/122.wav|12
1889 | ./Data/p230/130.wav|3
1890 | ./Data/p228/58.wav|1
1891 | ./Data/p227/38.wav|11
1892 | ./Data/p239/34.wav|7
1893 | ./Data/p240/137.wav|8
1894 | ./Data/p258/90.wav|16
1895 | ./Data/p258/138.wav|16
1896 | ./Data/p244/124.wav|9
1897 | ./Data/p239/167.wav|7
1898 | ./Data/p233/90.wav|5
1899 | ./Data/p239/172.wav|7
1900 | ./Data/p254/97.wav|14
1901 | ./Data/p259/29.wav|17
1902 | ./Data/p229/92.wav|2
1903 | ./Data/p227/11.wav|11
1904 | ./Data/p258/118.wav|16
1905 | ./Data/p244/69.wav|9
1906 | ./Data/p232/3.wav|12
1907 | ./Data/p256/28.wav|15
1908 | ./Data/p229/49.wav|2
1909 | ./Data/p236/82.wav|6
1910 | ./Data/p239/171.wav|7
1911 | ./Data/p254/127.wav|14
1912 | ./Data/p259/43.wav|17
1913 | ./Data/p228/21.wav|1
1914 | ./Data/p256/74.wav|15
1915 | ./Data/p226/76.wav|10
1916 | ./Data/p243/170.wav|13
1917 | ./Data/p239/39.wav|7
1918 | ./Data/p233/124.wav|5
1919 | ./Data/p229/13.wav|2
1920 | ./Data/p231/71.wav|4
1921 | ./Data/p229/118.wav|2
1922 | ./Data/p231/88.wav|4
1923 | ./Data/p231/55.wav|4
1924 | ./Data/p270/104.wav|18
1925 | ./Data/p270/110.wav|18
1926 | ./Data/p228/41.wav|1
1927 | ./Data/p258/2.wav|16
1928 | ./Data/p230/78.wav|3
1929 | ./Data/p231/80.wav|4
1930 | ./Data/p243/9.wav|13
1931 | ./Data/p239/16.wav|7
1932 | ./Data/p239/76.wav|7
1933 | ./Data/p226/126.wav|10
1934 | ./Data/p226/63.wav|10
1935 | ./Data/p233/46.wav|5
1936 | ./Data/p270/202.wav|18
1937 | ./Data/p239/164.wav|7
1938 | ./Data/p231/22.wav|4
1939 | ./Data/p259/24.wav|17
1940 | ./Data/p256/73.wav|15
1941 | ./Data/p259/10.wav|17
1942 | ./Data/p232/94.wav|12
1943 | ./Data/p273/30.wav|19
1944 | ./Data/p244/29.wav|9
1945 | ./Data/p226/129.wav|10
1946 | ./Data/p243/81.wav|13
1947 | ./Data/p236/121.wav|6
1948 | ./Data/p228/89.wav|1
1949 | ./Data/p231/81.wav|4
1950 | ./Data/p243/57.wav|13
1951 | ./Data/p236/40.wav|6
1952 | ./Data/p226/89.wav|10
1953 | ./Data/p244/44.wav|9
1954 | ./Data/p254/88.wav|14
1955 | ./Data/p227/108.wav|11
1956 | ./Data/p258/123.wav|16
1957 | ./Data/p233/95.wav|5
1958 | ./Data/p259/142.wav|17
1959 | ./Data/p231/73.wav|4
1960 | ./Data/p258/52.wav|16
1961 | ./Data/p236/89.wav|6
1962 | ./Data/p229/67.wav|2
1963 | ./Data/p258/46.wav|16
1964 | ./Data/p231/132.wav|4
1965 | ./Data/p227/41.wav|11
1966 | ./Data/p256/114.wav|15
1967 | ./Data/p232/10.wav|12
1968 | ./Data/p225/46.wav|0
1969 | ./Data/p231/61.wav|4
1970 | ./Data/p229/30.wav|2
1971 | ./Data/p236/101.wav|6
1972 | ./Data/p256/20.wav|15
1973 | ./Data/p226/60.wav|10
1974 | ./Data/p259/18.wav|17
1975 | ./Data/p236/151.wav|6
1976 | ./Data/p233/130.wav|5
1977 | ./Data/p273/91.wav|19
1978 | ./Data/p225/59.wav|0
1979 | ./Data/p227/83.wav|11
1980 | ./Data/p226/127.wav|10
1981 | ./Data/p270/137.wav|18
1982 | ./Data/p258/95.wav|16
1983 | ./Data/p227/42.wav|11
1984 | ./Data/p230/108.wav|3
1985 | ./Data/p243/137.wav|13
1986 | ./Data/p228/157.wav|1
1987 | ./Data/p243/105.wav|13
1988 | ./Data/p228/133.wav|1
1989 | ./Data/p270/93.wav|18
1990 | ./Data/p256/86.wav|15
1991 | ./Data/p254/17.wav|14
1992 | ./Data/p227/135.wav|11
1993 | ./Data/p228/118.wav|1
1994 | ./Data/p239/142.wav|7
1995 | ./Data/p273/137.wav|19
1996 | ./Data/p259/79.wav|17
1997 | ./Data/p259/108.wav|17
1998 | ./Data/p226/15.wav|10
1999 | ./Data/p231/43.wav|4
2000 | ./Data/p256/16.wav|15
2001 | ./Data/p232/20.wav|12
2002 | ./Data/p258/35.wav|16
2003 | ./Data/p243/141.wav|13
2004 | ./Data/p232/104.wav|12
2005 | ./Data/p259/58.wav|17
2006 | ./Data/p258/82.wav|16
2007 | ./Data/p233/76.wav|5
2008 | ./Data/p270/126.wav|18
2009 | ./Data/p236/70.wav|6
2010 | ./Data/p240/49.wav|8
2011 | ./Data/p256/106.wav|15
2012 | ./Data/p254/55.wav|14
2013 | ./Data/p270/2.wav|18
2014 | ./Data/p270/143.wav|18
2015 | ./Data/p229/48.wav|2
2016 | ./Data/p244/6.wav|9
2017 | ./Data/p233/65.wav|5
2018 | ./Data/p233/18.wav|5
2019 | ./Data/p244/87.wav|9
2020 | ./Data/p236/133.wav|6
2021 | ./Data/p227/2.wav|11
2022 | ./Data/p227/17.wav|11
2023 | ./Data/p273/111.wav|19
2024 | ./Data/p230/98.wav|3
2025 | ./Data/p226/120.wav|10
2026 | ./Data/p226/112.wav|10
2027 | ./Data/p230/161.wav|3
2028 | ./Data/p254/79.wav|14
2029 | ./Data/p230/101.wav|3
2030 | ./Data/p239/96.wav|7
2031 | ./Data/p228/159.wav|1
2032 | ./Data/p230/24.wav|3
2033 | ./Data/p240/28.wav|8
2034 | ./Data/p254/125.wav|14
2035 | ./Data/p259/168.wav|17
2036 | ./Data/p228/18.wav|1
2037 | ./Data/p270/88.wav|18
2038 | ./Data/p270/25.wav|18
2039 | ./Data/p231/89.wav|4
2040 | ./Data/p230/14.wav|3
2041 | ./Data/p254/63.wav|14
2042 | ./Data/p233/53.wav|5
2043 | ./Data/p225/54.wav|0
2044 | ./Data/p243/19.wav|13
2045 | ./Data/p259/139.wav|17
2046 | ./Data/p229/87.wav|2
2047 | ./Data/p232/56.wav|12
2048 | ./Data/p270/97.wav|18
2049 | ./Data/p232/95.wav|12
2050 | ./Data/p232/86.wav|12
2051 | ./Data/p259/137.wav|17
2052 | ./Data/p228/147.wav|1
2053 | ./Data/p273/112.wav|19
2054 | ./Data/p243/80.wav|13
2055 | ./Data/p233/72.wav|5
2056 | ./Data/p233/114.wav|5
2057 | ./Data/p240/23.wav|8
2058 | ./Data/p236/164.wav|6
2059 | ./Data/p236/144.wav|6
2060 | ./Data/p254/116.wav|14
2061 | ./Data/p273/105.wav|19
2062 | ./Data/p239/48.wav|7
2063 | ./Data/p236/68.wav|6
2064 | ./Data/p233/87.wav|5
2065 | ./Data/p239/50.wav|7
2066 | ./Data/p256/66.wav|15
2067 | ./Data/p270/159.wav|18
2068 | ./Data/p273/53.wav|19
2069 | ./Data/p254/28.wav|14
2070 | ./Data/p259/28.wav|17
2071 | ./Data/p227/89.wav|11
2072 | ./Data/p243/1.wav|13
2073 | ./Data/p239/61.wav|7
2074 | ./Data/p226/28.wav|10
2075 | ./Data/p232/113.wav|12
2076 | ./Data/p225/38.wav|0
2077 | ./Data/p236/128.wav|6
2078 | ./Data/p225/3.wav|0
2079 | ./Data/p258/83.wav|16
2080 | ./Data/p270/195.wav|18
2081 | ./Data/p231/69.wav|4
2082 | ./Data/p254/49.wav|14
2083 | ./Data/p226/135.wav|10
2084 | ./Data/p230/3.wav|3
2085 | ./Data/p228/124.wav|1
2086 | ./Data/p233/119.wav|5
2087 | ./Data/p229/31.wav|2
2088 | ./Data/p256/54.wav|15
2089 | ./Data/p258/121.wav|16
2090 | ./Data/p231/57.wav|4
2091 | ./Data/p244/84.wav|9
2092 | ./Data/p244/113.wav|9
2093 | ./Data/p228/71.wav|1
2094 | ./Data/p270/86.wav|18
2095 | ./Data/p254/98.wav|14
2096 | ./Data/p225/19.wav|0
2097 | ./Data/p258/21.wav|16
2098 | ./Data/p259/60.wav|17
2099 | ./Data/p227/105.wav|11
2100 | ./Data/p258/142.wav|16
2101 | ./Data/p230/52.wav|3
2102 | ./Data/p227/6.wav|11
2103 | ./Data/p244/139.wav|9
2104 | ./Data/p226/128.wav|10
2105 | ./Data/p239/70.wav|7
2106 | ./Data/p273/28.wav|19
2107 | ./Data/p230/171.wav|3
2108 | ./Data/p270/113.wav|18
2109 | ./Data/p259/19.wav|17
2110 | ./Data/p225/68.wav|0
2111 | ./Data/p239/73.wav|7
2112 | ./Data/p254/44.wav|14
2113 | ./Data/p240/113.wav|8
2114 | ./Data/p244/77.wav|9
2115 | ./Data/p259/49.wav|17
2116 | ./Data/p225/86.wav|0
2117 | ./Data/p258/94.wav|16
2118 | ./Data/p244/17.wav|9
2119 | ./Data/p227/12.wav|11
2120 | ./Data/p239/150.wav|7
2121 | ./Data/p225/10.wav|0
2122 | ./Data/p230/114.wav|3
2123 | ./Data/p258/69.wav|16
2124 | ./Data/p231/117.wav|4
2125 | ./Data/p244/23.wav|9
2126 | ./Data/p273/60.wav|19
2127 | ./Data/p259/156.wav|17
2128 | ./Data/p239/158.wav|7
2129 | ./Data/p244/102.wav|9
2130 | ./Data/p236/85.wav|6
2131 | ./Data/p259/2.wav|17
2132 | ./Data/p259/83.wav|17
2133 | ./Data/p226/40.wav|10
2134 | ./Data/p270/34.wav|18
2135 | ./Data/p240/99.wav|8
2136 | ./Data/p259/95.wav|17
2137 | ./Data/p240/79.wav|8
2138 | ./Data/p239/102.wav|7
2139 | ./Data/p273/57.wav|19
2140 | ./Data/p243/85.wav|13
2141 | ./Data/p239/149.wav|7
2142 | ./Data/p232/28.wav|12
2143 | ./Data/p254/25.wav|14
2144 | ./Data/p233/42.wav|5
2145 | ./Data/p227/39.wav|11
2146 | ./Data/p270/77.wav|18
2147 | ./Data/p233/51.wav|5
2148 | ./Data/p256/100.wav|15
2149 | ./Data/p258/140.wav|16
2150 | ./Data/p229/131.wav|2
2151 | ./Data/p243/52.wav|13
2152 | ./Data/p258/84.wav|16
2153 | ./Data/p229/138.wav|2
2154 | ./Data/p240/61.wav|8
2155 | ./Data/p254/27.wav|14
2156 | ./Data/p232/21.wav|12
2157 | ./Data/p226/38.wav|10
2158 | ./Data/p230/158.wav|3
2159 | ./Data/p256/52.wav|15
2160 | ./Data/p243/95.wav|13
2161 | ./Data/p243/89.wav|13
2162 | ./Data/p226/61.wav|10
2163 | ./Data/p230/117.wav|3
2164 | ./Data/p230/92.wav|3
2165 | ./Data/p236/55.wav|6
2166 | ./Data/p254/18.wav|14
2167 | ./Data/p254/129.wav|14
2168 | ./Data/p259/113.wav|17
2169 | ./Data/p225/25.wav|0
2170 | ./Data/p240/134.wav|8
2171 | ./Data/p230/86.wav|3
2172 | ./Data/p256/84.wav|15
2173 | ./Data/p228/99.wav|1
2174 | ./Data/p239/90.wav|7
2175 | ./Data/p230/155.wav|3
2176 | ./Data/p228/40.wav|1
2177 | ./Data/p254/72.wav|14
2178 | ./Data/p231/38.wav|4
2179 | ./Data/p225/32.wav|0
2180 | ./Data/p228/22.wav|1
2181 | ./Data/p231/7.wav|4
2182 | ./Data/p254/39.wav|14
2183 | ./Data/p240/112.wav|8
2184 | ./Data/p270/183.wav|18
2185 | ./Data/p270/60.wav|18
2186 | ./Data/p236/120.wav|6
2187 | ./Data/p239/145.wav|7
2188 | ./Data/p240/31.wav|8
2189 | ./Data/p229/115.wav|2
2190 | ./Data/p233/121.wav|5
2191 | ./Data/p228/33.wav|1
2192 | ./Data/p228/83.wav|1
2193 | ./Data/p258/58.wav|16
2194 | ./Data/p239/106.wav|7
2195 | ./Data/p273/123.wav|19
2196 | ./Data/p244/50.wav|9
2197 | ./Data/p229/50.wav|2
2198 | ./Data/p270/131.wav|18
2199 | ./Data/p236/8.wav|6
2200 | ./Data/p244/114.wav|9
2201 | ./Data/p230/153.wav|3
2202 | ./Data/p226/53.wav|10
2203 | ./Data/p240/93.wav|8
2204 | ./Data/p229/122.wav|2
2205 | ./Data/p256/90.wav|15
2206 | ./Data/p231/112.wav|4
2207 | ./Data/p270/48.wav|18
2208 | ./Data/p230/36.wav|3
2209 | ./Data/p230/135.wav|3
2210 | ./Data/p259/172.wav|17
2211 | ./Data/p229/55.wav|2
2212 | ./Data/p244/60.wav|9
2213 | ./Data/p232/75.wav|12
2214 | ./Data/p259/68.wav|17
2215 | ./Data/p233/7.wav|5
2216 | ./Data/p233/3.wav|5
2217 | ./Data/p226/141.wav|10
2218 | ./Data/p254/32.wav|14
2219 | ./Data/p239/26.wav|7
2220 | ./Data/p226/119.wav|10
2221 | ./Data/p239/173.wav|7
2222 | ./Data/p230/157.wav|3
2223 | ./Data/p236/157.wav|6
2224 | ./Data/p226/13.wav|10
2225 | ./Data/p254/68.wav|14
2226 | ./Data/p225/87.wav|0
2227 | ./Data/p231/118.wav|4
2228 | ./Data/p240/98.wav|8
2229 | ./Data/p233/5.wav|5
2230 | ./Data/p227/56.wav|11
2231 | ./Data/p239/93.wav|7
2232 | ./Data/p240/25.wav|8
2233 | ./Data/p243/142.wav|13
2234 | ./Data/p254/110.wav|14
2235 | ./Data/p230/138.wav|3
2236 | ./Data/p226/16.wav|10
2237 | ./Data/p270/189.wav|18
2238 | ./Data/p229/95.wav|2
2239 | ./Data/p231/37.wav|4
2240 | ./Data/p240/44.wav|8
2241 | ./Data/p228/46.wav|1
2242 | ./Data/p236/62.wav|6
2243 | ./Data/p226/20.wav|10
2244 | ./Data/p228/105.wav|1
2245 | ./Data/p258/44.wav|16
2246 | ./Data/p258/23.wav|16
2247 | ./Data/p270/108.wav|18
2248 | ./Data/p243/151.wav|13
2249 | ./Data/p239/170.wav|7
2250 | ./Data/p244/100.wav|9
2251 | ./Data/p258/81.wav|16
2252 | ./Data/p236/153.wav|6
2253 | ./Data/p229/5.wav|2
2254 | ./Data/p256/112.wav|15
2255 | ./Data/p258/70.wav|16
2256 | ./Data/p240/57.wav|8
2257 | ./Data/p244/36.wav|9
2258 | ./Data/p273/19.wav|19
2259 | ./Data/p233/75.wav|5
2260 | ./Data/p259/111.wav|17
2261 | ./Data/p243/100.wav|13
2262 | ./Data/p226/86.wav|10
2263 | ./Data/p256/26.wav|15
2264 | ./Data/p236/22.wav|6
2265 | ./Data/p229/124.wav|2
2266 | ./Data/p229/62.wav|2
2267 | ./Data/p258/87.wav|16
2268 | ./Data/p232/22.wav|12
2269 | ./Data/p259/158.wav|17
2270 | ./Data/p229/135.wav|2
2271 | ./Data/p233/118.wav|5
2272 | ./Data/p236/134.wav|6
2273 | ./Data/p226/34.wav|10
2274 | ./Data/p236/93.wav|6
2275 | ./Data/p243/108.wav|13
2276 | ./Data/p270/177.wav|18
2277 | ./Data/p239/30.wav|7
2278 | ./Data/p273/17.wav|19
2279 | ./Data/p231/110.wav|4
2280 | ./Data/p229/119.wav|2
2281 | ./Data/p243/130.wav|13
2282 | ./Data/p256/127.wav|15
2283 | ./Data/p226/105.wav|10
2284 | ./Data/p229/52.wav|2
2285 | ./Data/p226/54.wav|10
2286 | ./Data/p273/87.wav|19
2287 | ./Data/p270/57.wav|18
2288 | ./Data/p240/131.wav|8
2289 | ./Data/p273/117.wav|19
2290 | ./Data/p240/77.wav|8
2291 | ./Data/p233/32.wav|5
2292 | ./Data/p236/25.wav|6
2293 | ./Data/p227/79.wav|11
2294 | ./Data/p258/64.wav|16
2295 | ./Data/p240/92.wav|8
2296 | ./Data/p244/74.wav|9
2297 | ./Data/p228/120.wav|1
2298 | ./Data/p230/45.wav|3
2299 | ./Data/p225/89.wav|0
2300 | ./Data/p226/95.wav|10
2301 | ./Data/p270/80.wav|18
2302 | ./Data/p226/111.wav|10
2303 | ./Data/p243/2.wav|13
2304 | ./Data/p259/6.wav|17
2305 | ./Data/p227/85.wav|11
2306 | ./Data/p233/106.wav|5
2307 | ./Data/p227/14.wav|11
2308 | ./Data/p231/50.wav|4
2309 | ./Data/p230/139.wav|3
2310 | ./Data/p229/70.wav|2
2311 | ./Data/p258/14.wav|16
2312 | ./Data/p240/116.wav|8
2313 | ./Data/p225/64.wav|0
2314 | ./Data/p225/8.wav|0
2315 | ./Data/p243/113.wav|13
2316 | ./Data/p254/102.wav|14
2317 | ./Data/p270/148.wav|18
2318 | ./Data/p232/12.wav|12
2319 | ./Data/p259/22.wav|17
2320 | ./Data/p273/4.wav|19
2321 | ./Data/p244/133.wav|9
2322 | ./Data/p228/101.wav|1
2323 | ./Data/p273/31.wav|19
2324 | ./Data/p258/76.wav|16
2325 | ./Data/p227/146.wav|11
2326 | ./Data/p231/54.wav|4
2327 | ./Data/p236/37.wav|6
2328 | ./Data/p244/82.wav|9
2329 | ./Data/p225/17.wav|0
2330 | ./Data/p243/76.wav|13
2331 | ./Data/p273/140.wav|19
2332 | ./Data/p239/15.wav|7
2333 | ./Data/p230/19.wav|3
2334 | ./Data/p240/117.wav|8
2335 | ./Data/p244/94.wav|9
2336 | ./Data/p236/26.wav|6
2337 | ./Data/p259/99.wav|17
2338 | ./Data/p225/77.wav|0
2339 | ./Data/p244/31.wav|9
2340 | ./Data/p244/98.wav|9
2341 | ./Data/p243/59.wav|13
2342 | ./Data/p228/163.wav|1
2343 | ./Data/p270/141.wav|18
2344 | ./Data/p230/94.wav|3
2345 | ./Data/p228/110.wav|1
2346 | ./Data/p243/160.wav|13
2347 | ./Data/p239/162.wav|7
2348 | ./Data/p232/112.wav|12
2349 | ./Data/p273/54.wav|19
2350 | ./Data/p259/110.wav|17
2351 | ./Data/p244/64.wav|9
2352 | ./Data/p259/170.wav|17
2353 | ./Data/p230/53.wav|3
2354 | ./Data/p228/8.wav|1
2355 | ./Data/p232/80.wav|12
2356 | ./Data/p273/56.wav|19
2357 | ./Data/p256/93.wav|15
2358 | ./Data/p258/50.wav|16
2359 | ./Data/p231/41.wav|4
2360 | ./Data/p236/76.wav|6
2361 | ./Data/p229/65.wav|2
2362 | ./Data/p243/46.wav|13
2363 | ./Data/p228/31.wav|1
2364 | ./Data/p240/89.wav|8
2365 | ./Data/p240/119.wav|8
2366 | ./Data/p243/31.wav|13
2367 | ./Data/p273/122.wav|19
2368 | ./Data/p236/113.wav|6
2369 | ./Data/p232/67.wav|12
2370 | ./Data/p270/188.wav|18
2371 | ./Data/p256/46.wav|15
2372 | ./Data/p230/12.wav|3
2373 | ./Data/p236/156.wav|6
2374 | ./Data/p243/157.wav|13
2375 | ./Data/p239/22.wav|7
2376 | ./Data/p232/107.wav|12
2377 | ./Data/p229/28.wav|2
2378 | ./Data/p236/31.wav|6
2379 | ./Data/p254/50.wav|14
2380 | ./Data/p232/43.wav|12
2381 | ./Data/p244/142.wav|9
2382 | ./Data/p270/186.wav|18
2383 | ./Data/p258/139.wav|16
2384 | ./Data/p228/156.wav|1
2385 | ./Data/p256/12.wav|15
2386 | ./Data/p256/63.wav|15
2387 | ./Data/p230/116.wav|3
2388 | ./Data/p254/131.wav|14
2389 | ./Data/p243/139.wav|13
2390 | ./Data/p226/93.wav|10
2391 | ./Data/p239/98.wav|7
2392 | ./Data/p256/94.wav|15
2393 | ./Data/p243/102.wav|13
2394 | ./Data/p240/3.wav|8
2395 | ./Data/p258/86.wav|16
2396 | ./Data/p227/28.wav|11
2397 | ./Data/p228/49.wav|1
2398 | ./Data/p270/47.wav|18
2399 | ./Data/p226/88.wav|10
2400 | ./Data/p232/36.wav|12
2401 | ./Data/p259/90.wav|17
2402 | ./Data/p244/5.wav|9
2403 | ./Data/p243/122.wav|13
2404 | ./Data/p254/10.wav|14
2405 | ./Data/p254/64.wav|14
2406 | ./Data/p273/47.wav|19
2407 | ./Data/p243/171.wav|13
2408 | ./Data/p243/4.wav|13
2409 | ./Data/p230/128.wav|3
2410 | ./Data/p229/84.wav|2
2411 | ./Data/p259/39.wav|17
2412 | ./Data/p236/5.wav|6
2413 | ./Data/p225/47.wav|0
2414 | ./Data/p258/134.wav|16
2415 | ./Data/p259/12.wav|17
2416 | ./Data/p244/9.wav|9
2417 | ./Data/p227/98.wav|11
2418 | ./Data/p227/65.wav|11
2419 | ./Data/p226/114.wav|10
2420 | ./Data/p229/113.wav|2
2421 | ./Data/p240/16.wav|8
2422 | ./Data/p227/118.wav|11
2423 | ./Data/p258/56.wav|16
2424 | ./Data/p270/150.wav|18
2425 | ./Data/p256/85.wav|15
2426 | ./Data/p259/92.wav|17
2427 | ./Data/p239/84.wav|7
2428 | ./Data/p240/86.wav|8
2429 | ./Data/p225/11.wav|0
2430 | ./Data/p226/25.wav|10
2431 | ./Data/p270/65.wav|18
2432 | ./Data/p239/79.wav|7
2433 | ./Data/p240/76.wav|8
2434 | ./Data/p270/190.wav|18
2435 | ./Data/p236/163.wav|6
2436 | ./Data/p236/36.wav|6
2437 | ./Data/p240/41.wav|8
2438 | ./Data/p226/2.wav|10
2439 | ./Data/p230/104.wav|3
2440 | ./Data/p243/106.wav|13
2441 | ./Data/p243/90.wav|13
2442 | ./Data/p240/27.wav|8
2443 | ./Data/p240/30.wav|8
2444 | ./Data/p231/121.wav|4
2445 | ./Data/p239/8.wav|7
2446 | ./Data/p230/10.wav|3
2447 | ./Data/p239/104.wav|7
2448 | ./Data/p233/96.wav|5
2449 | ./Data/p236/33.wav|6
2450 | ./Data/p254/85.wav|14
2451 | ./Data/p227/26.wav|11
2452 | ./Data/p233/134.wav|5
2453 | ./Data/p230/48.wav|3
2454 | ./Data/p232/59.wav|12
2455 | ./Data/p239/156.wav|7
2456 | ./Data/p236/84.wav|6
2457 | ./Data/p228/63.wav|1
2458 | ./Data/p229/24.wav|2
2459 | ./Data/p236/155.wav|6
2460 | ./Data/p228/138.wav|1
2461 | ./Data/p270/185.wav|18
2462 | ./Data/p228/76.wav|1
2463 | ./Data/p254/115.wav|14
2464 | ./Data/p231/52.wav|4
2465 | ./Data/p273/8.wav|19
2466 | ./Data/p228/132.wav|1
2467 | ./Data/p273/59.wav|19
2468 | ./Data/p229/73.wav|2
2469 | ./Data/p259/152.wav|17
2470 | ./Data/p230/31.wav|3
2471 | ./Data/p230/35.wav|3
2472 | ./Data/p258/80.wav|16
2473 | ./Data/p225/61.wav|0
2474 | ./Data/p236/21.wav|6
2475 | ./Data/p232/127.wav|12
2476 | ./Data/p256/72.wav|15
2477 | ./Data/p244/123.wav|9
2478 | ./Data/p244/141.wav|9
2479 | ./Data/p270/69.wav|18
2480 | ./Data/p227/51.wav|11
2481 | ./Data/p273/11.wav|19
2482 | ./Data/p243/112.wav|13
2483 | ./Data/p254/16.wav|14
2484 | ./Data/p226/3.wav|10
2485 | ./Data/p231/36.wav|4
2486 | ./Data/p243/159.wav|13
2487 | ./Data/p228/55.wav|1
2488 | ./Data/p229/18.wav|2
2489 | ./Data/p273/22.wav|19
2490 | ./Data/p270/101.wav|18
2491 | ./Data/p227/62.wav|11
2492 | ./Data/p270/111.wav|18
2493 | ./Data/p254/73.wav|14
2494 | ./Data/p256/81.wav|15
2495 | ./Data/p226/116.wav|10
2496 | ./Data/p236/154.wav|6
2497 | ./Data/p233/98.wav|5
2498 | ./Data/p239/68.wav|7
2499 | ./Data/p273/69.wav|19
2500 | ./Data/p236/92.wav|6
2501 | ./Data/p273/81.wav|19
2502 | ./Data/p225/43.wav|0
2503 | ./Data/p230/27.wav|3
2504 | ./Data/p227/54.wav|11
2505 | ./Data/p233/113.wav|5
2506 | ./Data/p236/23.wav|6
2507 | ./Data/p236/51.wav|6
2508 | ./Data/p233/50.wav|5
2509 | ./Data/p225/76.wav|0
2510 | ./Data/p244/21.wav|9
2511 | ./Data/p228/53.wav|1
2512 | ./Data/p240/148.wav|8
2513 | ./Data/p243/173.wav|13
2514 | ./Data/p270/105.wav|18
2515 | ./Data/p227/13.wav|11
2516 | ./Data/p228/121.wav|1
2517 | ./Data/p233/128.wav|5
2518 | ./Data/p256/82.wav|15
2519 | ./Data/p244/76.wav|9
2520 | ./Data/p232/9.wav|12
2521 | ./Data/p239/4.wav|7
2522 | ./Data/p240/106.wav|8
2523 | ./Data/p270/81.wav|18
2524 | ./Data/p225/48.wav|0
2525 | ./Data/p254/67.wav|14
2526 | ./Data/p240/66.wav|8
2527 | ./Data/p259/47.wav|17
2528 | ./Data/p230/63.wav|3
2529 | ./Data/p230/141.wav|3
2530 | ./Data/p231/137.wav|4
2531 | ./Data/p227/133.wav|11
2532 | ./Data/p259/100.wav|17
2533 | ./Data/p259/171.wav|17
2534 | ./Data/p240/56.wav|8
2535 | ./Data/p273/126.wav|19
2536 | ./Data/p256/32.wav|15
2537 | ./Data/p270/79.wav|18
2538 | ./Data/p227/46.wav|11
2539 | ./Data/p228/51.wav|1
2540 | ./Data/p243/54.wav|13
2541 | ./Data/p258/141.wav|16
2542 | ./Data/p226/31.wav|10
2543 | ./Data/p236/137.wav|6
2544 | ./Data/p230/30.wav|3
2545 | ./Data/p236/34.wav|6
2546 | ./Data/p228/35.wav|1
2547 | ./Data/p244/56.wav|9
2548 | ./Data/p230/107.wav|3
2549 | ./Data/p240/36.wav|8
2550 | ./Data/p233/62.wav|5
2551 | ./Data/p239/112.wav|7
2552 | ./Data/p231/42.wav|4
2553 | ./Data/p256/9.wav|15
2554 | ./Data/p227/23.wav|11
2555 | ./Data/p236/32.wav|6
2556 | ./Data/p228/67.wav|1
2557 | ./Data/p225/72.wav|0
2558 | ./Data/p232/82.wav|12
2559 | ./Data/p244/68.wav|9
2560 | ./Data/p230/145.wav|3
2561 | ./Data/p239/5.wav|7
2562 | ./Data/p230/154.wav|3
2563 | ./Data/p232/98.wav|12
2564 | ./Data/p243/136.wav|13
2565 | ./Data/p228/115.wav|1
2566 | ./Data/p226/5.wav|10
2567 | ./Data/p240/52.wav|8
2568 | ./Data/p270/170.wav|18
2569 | ./Data/p243/93.wav|13
2570 | ./Data/p243/26.wav|13
2571 | ./Data/p230/136.wav|3
2572 | ./Data/p226/97.wav|10
2573 | ./Data/p229/136.wav|2
2574 | ./Data/p227/136.wav|11
2575 | ./Data/p236/119.wav|6
2576 | ./Data/p232/14.wav|12
2577 | ./Data/p254/138.wav|14
2578 | ./Data/p240/143.wav|8
2579 | ./Data/p259/122.wav|17
2580 | ./Data/p270/205.wav|18
2581 | ./Data/p254/100.wav|14
2582 | ./Data/p270/149.wav|18
2583 | ./Data/p259/9.wav|17
2584 | ./Data/p226/96.wav|10
2585 | ./Data/p230/23.wav|3
2586 | ./Data/p244/72.wav|9
2587 | ./Data/p259/73.wav|17
2588 | ./Data/p227/68.wav|11
2589 | ./Data/p226/75.wav|10
2590 | ./Data/p236/109.wav|6
2591 | ./Data/p258/102.wav|16
2592 | ./Data/p232/44.wav|12
2593 | ./Data/p243/27.wav|13
2594 | ./Data/p232/126.wav|12
2595 | ./Data/p240/14.wav|8
2596 | ./Data/p226/71.wav|10
2597 | ./Data/p230/88.wav|3
2598 | ./Data/p233/45.wav|5
2599 | ./Data/p244/103.wav|9
2600 | ./Data/p232/26.wav|12
2601 | ./Data/p229/101.wav|2
2602 | ./Data/p229/44.wav|2
2603 | ./Data/p232/123.wav|12
2604 | ./Data/p228/129.wav|1
2605 | ./Data/p273/32.wav|19
2606 | ./Data/p232/125.wav|12
2607 | ./Data/p240/103.wav|8
2608 | ./Data/p254/128.wav|14
2609 | ./Data/p254/34.wav|14
2610 | ./Data/p240/19.wav|8
2611 | ./Data/p232/89.wav|12
2612 | ./Data/p273/73.wav|19
2613 | ./Data/p231/109.wav|4
2614 | ./Data/p270/124.wav|18
2615 | ./Data/p244/112.wav|9
2616 | ./Data/p256/117.wav|15
2617 | ./Data/p244/88.wav|9
2618 | ./Data/p228/17.wav|1
2619 | ./Data/p233/86.wav|5
2620 | ./Data/p254/23.wav|14
2621 | ./Data/p233/59.wav|5
2622 | ./Data/p232/25.wav|12
2623 | ./Data/p231/108.wav|4
2624 | ./Data/p258/103.wav|16
2625 | ./Data/p232/69.wav|12
2626 | ./Data/p230/65.wav|3
2627 | ./Data/p240/73.wav|8
2628 | ./Data/p243/125.wav|13
2629 | ./Data/p256/92.wav|15
2630 | ./Data/p270/31.wav|18
2631 | ./Data/p256/44.wav|15
2632 | ./Data/p236/98.wav|6
2633 | ./Data/p228/90.wav|1
2634 | ./Data/p231/125.wav|4
2635 | ./Data/p232/64.wav|12
2636 | ./Data/p273/80.wav|19
2637 | ./Data/p227/32.wav|11
2638 | ./Data/p226/17.wav|10
2639 | ./Data/p226/69.wav|10
2640 | ./Data/p231/142.wav|4
2641 | ./Data/p225/65.wav|0
2642 | ./Data/p229/64.wav|2
2643 | ./Data/p240/70.wav|8
2644 | ./Data/p225/85.wav|0
2645 | ./Data/p259/166.wav|17
2646 | ./Data/p230/119.wav|3
2647 | ./Data/p258/135.wav|16
2648 | ./Data/p225/60.wav|0
2649 | ./Data/p239/74.wav|7
2650 | ./Data/p233/117.wav|5
2651 | ./Data/p226/44.wav|10
2652 | ./Data/p227/103.wav|11
2653 | ./Data/p228/45.wav|1
2654 | ./Data/p244/52.wav|9
2655 | ./Data/p230/168.wav|3
2656 | ./Data/p259/71.wav|17
2657 | ./Data/p270/109.wav|18
2658 | ./Data/p243/164.wav|13
2659 | ./Data/p243/36.wav|13
2660 | ./Data/p270/12.wav|18
2661 | ./Data/p229/125.wav|2
2662 | ./Data/p259/51.wav|17
2663 | ./Data/p225/81.wav|0
2664 | ./Data/p240/133.wav|8
2665 | ./Data/p270/130.wav|18
2666 | ./Data/p228/37.wav|1
2667 | ./Data/p228/39.wav|1
2668 | ./Data/p240/35.wav|8
2669 | ./Data/p231/124.wav|4
2670 | ./Data/p244/121.wav|9
2671 | ./Data/p270/133.wav|18
2672 | ./Data/p227/110.wav|11
2673 | ./Data/p244/134.wav|9
2674 | ./Data/p254/59.wav|14
2675 | ./Data/p239/35.wav|7
2676 | ./Data/p236/150.wav|6
2677 | ./Data/p227/40.wav|11
2678 | ./Data/p258/13.wav|16
2679 | ./Data/p240/123.wav|8
2680 | ./Data/p231/141.wav|4
2681 | ./Data/p228/151.wav|1
2682 | ./Data/p236/45.wav|6
2683 | ./Data/p273/5.wav|19
2684 | ./Data/p231/113.wav|4
2685 | ./Data/p256/103.wav|15
2686 | ./Data/p227/87.wav|11
2687 | ./Data/p270/173.wav|18
2688 | ./Data/p243/104.wav|13
2689 | ./Data/p240/141.wav|8
2690 | ./Data/p240/128.wav|8
2691 | ./Data/p259/50.wav|17
2692 | ./Data/p231/8.wav|4
2693 | ./Data/p226/82.wav|10
2694 | ./Data/p243/110.wav|13
2695 | ./Data/p243/101.wav|13
2696 | ./Data/p259/132.wav|17
2697 | ./Data/p227/99.wav|11
2698 | ./Data/p259/42.wav|17
2699 | ./Data/p229/29.wav|2
2700 | ./Data/p236/104.wav|6
2701 | ./Data/p259/34.wav|17
2702 | ./Data/p254/117.wav|14
2703 | ./Data/p227/29.wav|11
2704 | ./Data/p258/111.wav|16
2705 | ./Data/p229/9.wav|2
2706 | ./Data/p240/26.wav|8
2707 | ./Data/p259/89.wav|17
2708 | ./Data/p270/21.wav|18
2709 | ./Data/p254/101.wav|14
2710 | ./Data/p259/40.wav|17
2711 | ./Data/p240/7.wav|8
2712 | ./Data/p240/114.wav|8
2713 | ./Data/p230/176.wav|3
2714 | ./Data/p231/47.wav|4
2715 | ./Data/p239/37.wav|7
2716 | ./Data/p232/51.wav|12
2717 | ./Data/p270/142.wav|18
2718 | ./Data/p254/6.wav|14
2719 | ./Data/p225/50.wav|0
2720 | ./Data/p227/91.wav|11
2721 | ./Data/p259/149.wav|17
2722 | ./Data/p259/125.wav|17
2723 | ./Data/p229/107.wav|2
2724 | ./Data/p228/10.wav|1
2725 | ./Data/p231/107.wav|4
2726 | 


--------------------------------------------------------------------------------
/Data/val_list.txt:
--------------------------------------------------------------------------------
  1 | ./Data/p270/13.wav|18
  2 | ./Data/p273/94.wav|19
  3 | ./Data/p229/97.wav|2
  4 | ./Data/p232/117.wav|12
  5 | ./Data/p226/55.wav|10
  6 | ./Data/p259/102.wav|17
  7 | ./Data/p226/7.wav|10
  8 | ./Data/p254/26.wav|14
  9 | ./Data/p239/115.wav|7
 10 | ./Data/p239/86.wav|7
 11 | ./Data/p229/106.wav|2
 12 | ./Data/p244/43.wav|9
 13 | ./Data/p270/179.wav|18
 14 | ./Data/p273/6.wav|19
 15 | ./Data/p258/101.wav|16
 16 | ./Data/p273/62.wav|19
 17 | ./Data/p228/11.wav|1
 18 | ./Data/p273/103.wav|19
 19 | ./Data/p230/49.wav|3
 20 | ./Data/p233/23.wav|5
 21 | ./Data/p230/122.wav|3
 22 | ./Data/p239/80.wav|7
 23 | ./Data/p226/4.wav|10
 24 | ./Data/p240/50.wav|8
 25 | ./Data/p243/28.wav|13
 26 | ./Data/p236/95.wav|6
 27 | ./Data/p244/126.wav|9
 28 | ./Data/p244/40.wav|9
 29 | ./Data/p239/108.wav|7
 30 | ./Data/p273/72.wav|19
 31 | ./Data/p254/92.wav|14
 32 | ./Data/p231/116.wav|4
 33 | ./Data/p231/32.wav|4
 34 | ./Data/p243/117.wav|13
 35 | ./Data/p256/121.wav|15
 36 | ./Data/p243/3.wav|13
 37 | ./Data/p226/91.wav|10
 38 | ./Data/p256/53.wav|15
 39 | ./Data/p254/75.wav|14
 40 | ./Data/p243/150.wav|13
 41 | ./Data/p231/95.wav|4
 42 | ./Data/p228/81.wav|1
 43 | ./Data/p226/33.wav|10
 44 | ./Data/p232/71.wav|12
 45 | ./Data/p236/4.wav|6
 46 | ./Data/p236/132.wav|6
 47 | ./Data/p254/119.wav|14
 48 | ./Data/p236/7.wav|6
 49 | ./Data/p227/104.wav|11
 50 | ./Data/p226/59.wav|10
 51 | ./Data/p233/35.wav|5
 52 | ./Data/p231/23.wav|4
 53 | ./Data/p273/71.wav|19
 54 | ./Data/p240/74.wav|8
 55 | ./Data/p259/33.wav|17
 56 | ./Data/p259/118.wav|17
 57 | ./Data/p273/15.wav|19
 58 | ./Data/p226/115.wav|10
 59 | ./Data/p236/19.wav|6
 60 | ./Data/p226/57.wav|10
 61 | ./Data/p229/14.wav|2
 62 | ./Data/p243/98.wav|13
 63 | ./Data/p243/79.wav|13
 64 | ./Data/p231/12.wav|4
 65 | ./Data/p230/170.wav|3
 66 | ./Data/p228/114.wav|1
 67 | ./Data/p254/103.wav|14
 68 | ./Data/p256/108.wav|15
 69 | ./Data/p256/58.wav|15
 70 | ./Data/p229/23.wav|2
 71 | ./Data/p270/151.wav|18
 72 | ./Data/p259/36.wav|17
 73 | ./Data/p230/64.wav|3
 74 | ./Data/p226/134.wav|10
 75 | ./Data/p230/84.wav|3
 76 | ./Data/p270/91.wav|18
 77 | ./Data/p230/160.wav|3
 78 | ./Data/p236/15.wav|6
 79 | ./Data/p225/45.wav|0
 80 | ./Data/p239/62.wav|7
 81 | ./Data/p256/107.wav|15
 82 | ./Data/p258/144.wav|16
 83 | ./Data/p229/37.wav|2
 84 | ./Data/p226/108.wav|10
 85 | ./Data/p225/92.wav|0
 86 | ./Data/p227/138.wav|11
 87 | ./Data/p230/151.wav|3
 88 | ./Data/p229/90.wav|2
 89 | ./Data/p244/131.wav|9
 90 | ./Data/p231/1.wav|4
 91 | ./Data/p243/40.wav|13
 92 | ./Data/p226/131.wav|10
 93 | ./Data/p226/121.wav|10
 94 | ./Data/p270/119.wav|18
 95 | ./Data/p225/4.wav|0
 96 | ./Data/p243/39.wav|13
 97 | ./Data/p233/1.wav|5
 98 | ./Data/p239/117.wav|7
 99 | ./Data/p259/101.wav|17
100 | ./Data/p228/73.wav|1
101 | ./Data/p273/78.wav|19
102 | ./Data/p256/22.wav|15
103 | ./Data/p244/65.wav|9
104 | ./Data/p240/17.wav|8
105 | ./Data/p258/47.wav|16
106 | ./Data/p239/95.wav|7
107 | ./Data/p243/119.wav|13
108 | ./Data/p259/106.wav|17
109 | ./Data/p233/22.wav|5
110 | ./Data/p232/60.wav|12
111 | ./Data/p270/55.wav|18
112 | ./Data/p230/87.wav|3
113 | ./Data/p270/139.wav|18
114 | ./Data/p225/5.wav|0
115 | ./Data/p243/128.wav|13
116 | ./Data/p258/10.wav|16
117 | ./Data/p230/100.wav|3
118 | ./Data/p239/43.wav|7
119 | ./Data/p232/57.wav|12
120 | ./Data/p256/27.wav|15
121 | ./Data/p232/130.wav|12
122 | ./Data/p243/153.wav|13
123 | ./Data/p258/92.wav|16
124 | ./Data/p232/81.wav|12
125 | ./Data/p256/65.wav|15
126 | ./Data/p259/107.wav|17
127 | ./Data/p239/10.wav|7
128 | ./Data/p233/4.wav|5
129 | ./Data/p259/165.wav|17
130 | ./Data/p225/41.wav|0
131 | ./Data/p229/61.wav|2
132 | ./Data/p227/36.wav|11
133 | ./Data/p243/62.wav|13
134 | ./Data/p259/31.wav|17
135 | ./Data/p231/75.wav|4
136 | ./Data/p233/31.wav|5
137 | ./Data/p273/66.wav|19
138 | ./Data/p226/6.wav|10
139 | ./Data/p243/162.wav|13
140 | ./Data/p229/21.wav|2
141 | ./Data/p230/11.wav|3
142 | ./Data/p231/84.wav|4
143 | ./Data/p273/118.wav|19
144 | ./Data/p227/92.wav|11
145 | ./Data/p256/110.wav|15
146 | ./Data/p230/105.wav|3
147 | ./Data/p239/75.wav|7
148 | ./Data/p229/78.wav|2
149 | ./Data/p254/111.wav|14
150 | ./Data/p232/24.wav|12
151 | ./Data/p233/19.wav|5
152 | ./Data/p233/52.wav|5
153 | ./Data/p258/143.wav|16
154 | ./Data/p254/135.wav|14
155 | ./Data/p232/37.wav|12
156 | ./Data/p244/81.wav|9
157 | ./Data/p270/161.wav|18
158 | ./Data/p233/43.wav|5
159 | ./Data/p240/40.wav|8
160 | ./Data/p244/70.wav|9
161 | ./Data/p254/1.wav|14
162 | ./Data/p229/96.wav|2
163 | ./Data/p243/99.wav|13
164 | ./Data/p259/20.wav|17
165 | ./Data/p233/66.wav|5
166 | ./Data/p239/88.wav|7
167 | ./Data/p225/71.wav|0
168 | ./Data/p227/143.wav|11
169 | ./Data/p228/142.wav|1
170 | ./Data/p231/135.wav|4
171 | ./Data/p254/107.wav|14
172 | ./Data/p233/36.wav|5
173 | ./Data/p232/19.wav|12
174 | ./Data/p258/113.wav|16
175 | ./Data/p243/96.wav|13
176 | ./Data/p273/90.wav|19
177 | ./Data/p225/13.wav|0
178 | ./Data/p228/32.wav|1
179 | ./Data/p229/60.wav|2
180 | ./Data/p273/14.wav|19
181 | ./Data/p239/25.wav|7
182 | ./Data/p256/31.wav|15
183 | ./Data/p225/40.wav|0
184 | ./Data/p273/43.wav|19
185 | ./Data/p270/206.wav|18
186 | ./Data/p244/19.wav|9
187 | ./Data/p244/83.wav|9
188 | ./Data/p259/134.wav|17
189 | ./Data/p244/91.wav|9
190 | ./Data/p225/80.wav|0
191 | ./Data/p227/60.wav|11
192 | ./Data/p244/128.wav|9
193 | ./Data/p256/80.wav|15
194 | ./Data/p256/15.wav|15
195 | ./Data/p244/34.wav|9
196 | ./Data/p256/69.wav|15
197 | ./Data/p228/15.wav|1
198 | ./Data/p232/65.wav|12
199 | ./Data/p273/65.wav|19
200 | ./Data/p239/124.wav|7
201 | ./Data/p259/15.wav|17
202 | ./Data/p226/137.wav|10
203 | ./Data/p243/75.wav|13
204 | ./Data/p258/16.wav|16
205 | ./Data/p232/6.wav|12
206 | ./Data/p231/106.wav|4
207 | ./Data/p228/6.wav|1
208 | ./Data/p243/172.wav|13
209 | ./Data/p236/77.wav|6
210 | ./Data/p256/95.wav|15
211 | ./Data/p256/76.wav|15
212 | ./Data/p239/119.wav|7
213 | ./Data/p236/108.wav|6
214 | ./Data/p243/92.wav|13
215 | ./Data/p232/129.wav|12
216 | ./Data/p230/124.wav|3
217 | ./Data/p228/9.wav|1
218 | ./Data/p232/100.wav|12
219 | ./Data/p254/5.wav|14
220 | ./Data/p273/1.wav|19
221 | ./Data/p236/47.wav|6
222 | ./Data/p240/87.wav|8
223 | ./Data/p229/127.wav|2
224 | ./Data/p228/152.wav|1
225 | ./Data/p225/24.wav|0
226 | ./Data/p229/20.wav|2
227 | ./Data/p233/12.wav|5
228 | ./Data/p259/46.wav|17
229 | ./Data/p231/72.wav|4
230 | ./Data/p254/65.wav|14
231 | ./Data/p231/18.wav|4
232 | ./Data/p270/66.wav|18
233 | ./Data/p233/44.wav|5
234 | ./Data/p233/126.wav|5
235 | ./Data/p233/58.wav|5
236 | ./Data/p273/142.wav|19
237 | ./Data/p228/26.wav|1
238 | ./Data/p230/106.wav|3
239 | ./Data/p228/109.wav|1
240 | ./Data/p232/76.wav|12
241 | ./Data/p226/37.wav|10
242 | ./Data/p226/66.wav|10
243 | ./Data/p270/75.wav|18
244 | ./Data/p229/4.wav|2
245 | ./Data/p239/166.wav|7
246 | ./Data/p228/79.wav|1
247 | ./Data/p230/43.wav|3
248 | ./Data/p258/100.wav|16
249 | ./Data/p244/93.wav|9
250 | ./Data/p256/105.wav|15
251 | ./Data/p236/12.wav|6
252 | ./Data/p270/154.wav|18
253 | ./Data/p244/75.wav|9
254 | ./Data/p239/160.wav|7
255 | ./Data/p239/174.wav|7
256 | ./Data/p225/26.wav|0
257 | ./Data/p232/49.wav|12
258 | ./Data/p258/19.wav|16
259 | ./Data/p273/13.wav|19
260 | ./Data/p232/32.wav|12
261 | ./Data/p270/42.wav|18
262 | ./Data/p270/194.wav|18
263 | ./Data/p259/174.wav|17
264 | ./Data/p236/53.wav|6
265 | ./Data/p232/77.wav|12
266 | ./Data/p240/118.wav|8
267 | ./Data/p239/175.wav|7
268 | ./Data/p225/58.wav|0
269 | ./Data/p232/1.wav|12
270 | ./Data/p243/5.wav|13
271 | ./Data/p229/41.wav|2
272 | ./Data/p233/60.wav|5
273 | ./Data/p236/138.wav|6
274 | ./Data/p258/54.wav|16
275 | ./Data/p254/22.wav|14
276 | ./Data/p254/76.wav|14
277 | ./Data/p228/25.wav|1
278 | ./Data/p259/61.wav|17
279 | ./Data/p270/135.wav|18
280 | ./Data/p231/136.wav|4
281 | ./Data/p232/105.wav|12
282 | ./Data/p259/35.wav|17
283 | ./Data/p244/57.wav|9
284 | ./Data/p226/104.wav|10
285 | ./Data/p258/48.wav|16
286 | ./Data/p229/139.wav|2
287 | ./Data/p239/65.wav|7
288 | ./Data/p228/74.wav|1
289 | ./Data/p233/25.wav|5
290 | ./Data/p243/16.wav|13
291 | ./Data/p243/165.wav|13
292 | ./Data/p229/46.wav|2
293 | ./Data/p226/41.wav|10
294 | ./Data/p228/160.wav|1
295 | ./Data/p230/90.wav|3
296 | ./Data/p270/184.wav|18
297 | ./Data/p259/55.wav|17
298 | ./Data/p232/31.wav|12
299 | ./Data/p231/78.wav|4
300 | ./Data/p259/78.wav|17
301 | ./Data/p273/33.wav|19
302 | ./Data/p256/40.wav|15
303 | ./Data/p258/116.wav|16
304 | 


--------------------------------------------------------------------------------
/Demo/VCTK-corpus/p228/p228_023.wav:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yl4579/StarGANv2-VC/ebe97a0c41ea295dcea8f5e3e2b7e934edd01608/Demo/VCTK-corpus/p228/p228_023.wav


--------------------------------------------------------------------------------
/Demo/VCTK-corpus/p230/p230_023.wav:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yl4579/StarGANv2-VC/ebe97a0c41ea295dcea8f5e3e2b7e934edd01608/Demo/VCTK-corpus/p230/p230_023.wav


--------------------------------------------------------------------------------
/Demo/VCTK-corpus/p233/p233_023.wav:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yl4579/StarGANv2-VC/ebe97a0c41ea295dcea8f5e3e2b7e934edd01608/Demo/VCTK-corpus/p233/p233_023.wav


--------------------------------------------------------------------------------
/Demo/VCTK-corpus/p236/p236_023.wav:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yl4579/StarGANv2-VC/ebe97a0c41ea295dcea8f5e3e2b7e934edd01608/Demo/VCTK-corpus/p236/p236_023.wav


--------------------------------------------------------------------------------
/Demo/VCTK-corpus/p243/p243_023.wav:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yl4579/StarGANv2-VC/ebe97a0c41ea295dcea8f5e3e2b7e934edd01608/Demo/VCTK-corpus/p243/p243_023.wav


--------------------------------------------------------------------------------
/Demo/VCTK-corpus/p244/p244_023.wav:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yl4579/StarGANv2-VC/ebe97a0c41ea295dcea8f5e3e2b7e934edd01608/Demo/VCTK-corpus/p244/p244_023.wav


--------------------------------------------------------------------------------
/Demo/VCTK-corpus/p254/p254_023.wav:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yl4579/StarGANv2-VC/ebe97a0c41ea295dcea8f5e3e2b7e934edd01608/Demo/VCTK-corpus/p254/p254_023.wav


--------------------------------------------------------------------------------
/Demo/VCTK-corpus/p258/p258_023.wav:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yl4579/StarGANv2-VC/ebe97a0c41ea295dcea8f5e3e2b7e934edd01608/Demo/VCTK-corpus/p258/p258_023.wav


--------------------------------------------------------------------------------
/Demo/VCTK-corpus/p259/p259_023.wav:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yl4579/StarGANv2-VC/ebe97a0c41ea295dcea8f5e3e2b7e934edd01608/Demo/VCTK-corpus/p259/p259_023.wav


--------------------------------------------------------------------------------
/Demo/VCTK-corpus/p273/p273_023.wav:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yl4579/StarGANv2-VC/ebe97a0c41ea295dcea8f5e3e2b7e934edd01608/Demo/VCTK-corpus/p273/p273_023.wav


--------------------------------------------------------------------------------
/Demo/inference.ipynb:
--------------------------------------------------------------------------------
  1 | {
  2 |  "cells": [
  3 |   {
  4 |    "cell_type": "markdown",
  5 |    "metadata": {
  6 |     "id": "HwaQq4GRU_Nw"
  7 |    },
  8 |    "source": [
  9 |     "# StarGANv2-VC Demo (VCTK 20 Speakers)"
 10 |    ]
 11 |   },
 12 |   {
 13 |    "cell_type": "markdown",
 14 |    "metadata": {
 15 |     "id": "hCpoXuZeGKAn"
 16 |    },
 17 |    "source": [
 18 |     "### Utils"
 19 |    ]
 20 |   },
 21 |   {
 22 |    "cell_type": "code",
 23 |    "execution_count": null,
 24 |    "metadata": {},
 25 |    "outputs": [],
 26 |    "source": [
 27 |     "%cd .."
 28 |    ]
 29 |   },
 30 |   {
 31 |    "cell_type": "code",
 32 |    "execution_count": null,
 33 |    "metadata": {
 34 |     "colab": {
 35 |      "base_uri": "https://localhost:8080/"
 36 |     },
 37 |     "executionInfo": {
 38 |      "elapsed": 24923,
 39 |      "status": "ok",
 40 |      "timestamp": 1613984920200,
 41 |      "user": {
 42 |       "displayName": "Yinghao Li",
 43 |       "photoUrl": "",
 44 |       "userId": "12798981472803960591"
 45 |      },
 46 |      "user_tz": 300
 47 |     },
 48 |     "id": "3on9IjGhVGTP",
 49 |     "outputId": "63a799f8-564d-48c2-fb0f-e66c0cd9fdb8"
 50 |    },
 51 |    "outputs": [],
 52 |    "source": [
 53 |     "# load packages\n",
 54 |     "import random\n",
 55 |     "import yaml\n",
 56 |     "from munch import Munch\n",
 57 |     "import numpy as np\n",
 58 |     "import torch\n",
 59 |     "from torch import nn\n",
 60 |     "import torch.nn.functional as F\n",
 61 |     "import torchaudio\n",
 62 |     "import librosa\n",
 63 |     "\n",
 64 |     "from Utils.ASR.models import ASRCNN\n",
 65 |     "from Utils.JDC.model import JDCNet\n",
 66 |     "from models import Generator, MappingNetwork, StyleEncoder\n",
 67 |     "\n",
 68 |     "%matplotlib inline"
 69 |    ]
 70 |   },
 71 |   {
 72 |    "cell_type": "code",
 73 |    "execution_count": null,
 74 |    "metadata": {},
 75 |    "outputs": [],
 76 |    "source": [
 77 |     "# Source: http://speech.ee.ntu.edu.tw/~jjery2243542/resource/model/is18/en_speaker_used.txt\n",
 78 |     "# Source: https://github.com/jjery2243542/voice_conversion\n",
 79 |     "\n",
 80 |     "speakers = [225,228,229,230,231,233,236,239,240,244,226,227,232,243,254,256,258,259,270,273]\n",
 81 |     "\n",
 82 |     "to_mel = torchaudio.transforms.MelSpectrogram(\n",
 83 |     "    n_mels=80, n_fft=2048, win_length=1200, hop_length=300)\n",
 84 |     "mean, std = -4, 4\n",
 85 |     "\n",
 86 |     "def preprocess(wave):\n",
 87 |     "    wave_tensor = torch.from_numpy(wave).float()\n",
 88 |     "    mel_tensor = to_mel(wave_tensor)\n",
 89 |     "    mel_tensor = (torch.log(1e-5 + mel_tensor.unsqueeze(0)) - mean) / std\n",
 90 |     "    return mel_tensor\n",
 91 |     "\n",
 92 |     "def build_model(model_params={}):\n",
 93 |     "    args = Munch(model_params)\n",
 94 |     "    generator = Generator(args.dim_in, args.style_dim, args.max_conv_dim, w_hpf=args.w_hpf, F0_channel=args.F0_channel)\n",
 95 |     "    mapping_network = MappingNetwork(args.latent_dim, args.style_dim, args.num_domains, hidden_dim=args.max_conv_dim)\n",
 96 |     "    style_encoder = StyleEncoder(args.dim_in, args.style_dim, args.num_domains, args.max_conv_dim)\n",
 97 |     "    \n",
 98 |     "    nets_ema = Munch(generator=generator,\n",
 99 |     "                     mapping_network=mapping_network,\n",
100 |     "                     style_encoder=style_encoder)\n",
101 |     "\n",
102 |     "    return nets_ema\n",
103 |     "\n",
104 |     "def compute_style(speaker_dicts):\n",
105 |     "    reference_embeddings = {}\n",
106 |     "    for key, (path, speaker) in speaker_dicts.items():\n",
107 |     "        if path == \"\":\n",
108 |     "            label = torch.LongTensor([speaker]).to('cuda')\n",
109 |     "            latent_dim = starganv2.mapping_network.shared[0].in_features\n",
110 |     "            ref = starganv2.mapping_network(torch.randn(1, latent_dim).to('cuda'), label)\n",
111 |     "        else:\n",
112 |     "            wave, sr = librosa.load(path, sr=24000)\n",
113 |     "            audio, index = librosa.effects.trim(wave, top_db=30)\n",
114 |     "            if sr != 24000:\n",
115 |     "                wave = librosa.resample(wave, sr, 24000)\n",
116 |     "            mel_tensor = preprocess(wave).to('cuda')\n",
117 |     "\n",
118 |     "            with torch.no_grad():\n",
119 |     "                label = torch.LongTensor([speaker])\n",
120 |     "                ref = starganv2.style_encoder(mel_tensor.unsqueeze(1), label)\n",
121 |     "        reference_embeddings[key] = (ref, label)\n",
122 |     "    \n",
123 |     "    return reference_embeddings"
124 |    ]
125 |   },
126 |   {
127 |    "cell_type": "markdown",
128 |    "metadata": {},
129 |    "source": [
130 |     "### Load models"
131 |    ]
132 |   },
133 |   {
134 |    "cell_type": "code",
135 |    "execution_count": null,
136 |    "metadata": {},
137 |    "outputs": [],
138 |    "source": [
139 |     "# load F0 model\n",
140 |     "\n",
141 |     "F0_model = JDCNet(num_class=1, seq_len=192)\n",
142 |     "params = torch.load(\"Utils/JDC/bst.t7\")['net']\n",
143 |     "F0_model.load_state_dict(params)\n",
144 |     "_ = F0_model.eval()\n",
145 |     "F0_model = F0_model.to('cuda')"
146 |    ]
147 |   },
148 |   {
149 |    "cell_type": "code",
150 |    "execution_count": null,
151 |    "metadata": {
152 |     "executionInfo": {
153 |      "elapsed": 43003,
154 |      "status": "ok",
155 |      "timestamp": 1613984938321,
156 |      "user": {
157 |       "displayName": "Yinghao Li",
158 |       "photoUrl": "",
159 |       "userId": "12798981472803960591"
160 |      },
161 |      "user_tz": 300
162 |     },
163 |     "id": "NZA3ot-oF5t-"
164 |    },
165 |    "outputs": [],
166 |    "source": [
167 |     "# load vocoder\n",
168 |     "from parallel_wavegan.utils import load_model\n",
169 |     "vocoder = load_model(\"Vocoder/checkpoint-400000steps.pkl\").to('cuda').eval()\n",
170 |     "vocoder.remove_weight_norm()\n",
171 |     "_ = vocoder.eval()"
172 |    ]
173 |   },
174 |   {
175 |    "cell_type": "code",
176 |    "execution_count": null,
177 |    "metadata": {
178 |     "colab": {
179 |      "base_uri": "https://localhost:8080/"
180 |     },
181 |     "executionInfo": {
182 |      "elapsed": 24462,
183 |      "status": "ok",
184 |      "timestamp": 1613985522414,
185 |      "user": {
186 |       "displayName": "Yinghao Li",
187 |       "photoUrl": "",
188 |       "userId": "12798981472803960591"
189 |      },
190 |      "user_tz": 300
191 |     },
192 |     "id": "Ou4367LCyefA",
193 |     "outputId": "19c61f6f-f39a-43b9-9275-09418c2aebb4"
194 |    },
195 |    "outputs": [],
196 |    "source": [
197 |     "# load starganv2\n",
198 |     "\n",
199 |     "model_path = 'Models/VCTK20/epoch_00150.pth'\n",
200 |     "\n",
201 |     "with open('Models/VCTK20/config.yml') as f:\n",
202 |     "    starganv2_config = yaml.safe_load(f)\n",
203 |     "starganv2 = build_model(model_params=starganv2_config[\"model_params\"])\n",
204 |     "params = torch.load(model_path, map_location='cpu')\n",
205 |     "params = params['model_ema']\n",
206 |     "_ = [starganv2[key].load_state_dict(params[key]) for key in starganv2]\n",
207 |     "_ = [starganv2[key].eval() for key in starganv2]\n",
208 |     "starganv2.style_encoder = starganv2.style_encoder.to('cuda')\n",
209 |     "starganv2.mapping_network = starganv2.mapping_network.to('cuda')\n",
210 |     "starganv2.generator = starganv2.generator.to('cuda')"
211 |    ]
212 |   },
213 |   {
214 |    "cell_type": "markdown",
215 |    "metadata": {},
216 |    "source": [
217 |     "### Conversion"
218 |    ]
219 |   },
220 |   {
221 |    "cell_type": "code",
222 |    "execution_count": null,
223 |    "metadata": {},
224 |    "outputs": [],
225 |    "source": [
226 |     "# load input wave\n",
227 |     "selected_speakers = [273, 259, 258, 243, 254, 244, 236, 233, 230, 228]\n",
228 |     "k = random.choice(selected_speakers)\n",
229 |     "wav_path = 'Demo/VCTK-corpus/p' + str(k) + '/p' + str(k) + '_023.wav'\n",
230 |     "audio, source_sr = librosa.load(wav_path, sr=24000)\n",
231 |     "audio = audio / np.max(np.abs(audio))\n",
232 |     "audio.dtype = np.float32"
233 |    ]
234 |   },
235 |   {
236 |    "cell_type": "markdown",
237 |    "metadata": {},
238 |    "source": [
239 |     "#### Convert by style encoder"
240 |    ]
241 |   },
242 |   {
243 |    "cell_type": "code",
244 |    "execution_count": null,
245 |    "metadata": {},
246 |    "outputs": [],
247 |    "source": [
248 |     "# with reference, using style encoder\n",
249 |     "speaker_dicts = {}\n",
250 |     "for s in selected_speakers:\n",
251 |     "    k = s\n",
252 |     "    speaker_dicts['p' + str(s)] = ('Demo/VCTK-corpus/p' + str(k) + '/p' + str(k) + '_023.wav', speakers.index(s))\n",
253 |     "\n",
254 |     "reference_embeddings = compute_style(speaker_dicts)"
255 |    ]
256 |   },
257 |   {
258 |    "cell_type": "code",
259 |    "execution_count": null,
260 |    "metadata": {
261 |     "colab": {
262 |      "base_uri": "https://localhost:8080/",
263 |      "height": 333
264 |     },
265 |     "executionInfo": {
266 |      "elapsed": 1424,
267 |      "status": "ok",
268 |      "timestamp": 1613986299525,
269 |      "user": {
270 |       "displayName": "Yinghao Li",
271 |       "photoUrl": "",
272 |       "userId": "12798981472803960591"
273 |      },
274 |      "user_tz": 300
275 |     },
276 |     "id": "T5tahObUyN-d",
277 |     "outputId": "f4f38742-2235-4f59-cb2a-5008912cd870",
278 |     "scrolled": true
279 |    },
280 |    "outputs": [],
281 |    "source": [
282 |     "# conversion \n",
283 |     "import time\n",
284 |     "start = time.time()\n",
285 |     "    \n",
286 |     "source = preprocess(audio).to('cuda:0')\n",
287 |     "keys = []\n",
288 |     "converted_samples = {}\n",
289 |     "reconstructed_samples = {}\n",
290 |     "converted_mels = {}\n",
291 |     "\n",
292 |     "for key, (ref, _) in reference_embeddings.items():\n",
293 |     "    with torch.no_grad():\n",
294 |     "        f0_feat = F0_model.get_feature_GAN(source.unsqueeze(1))\n",
295 |     "        out = starganv2.generator(source.unsqueeze(1), ref, F0=f0_feat)\n",
296 |     "        \n",
297 |     "        c = out.transpose(-1, -2).squeeze().to('cuda')\n",
298 |     "        y_out = vocoder.inference(c)\n",
299 |     "        y_out = y_out.view(-1).cpu()\n",
300 |     "\n",
301 |     "        if key not in speaker_dicts or speaker_dicts[key][0] == \"\":\n",
302 |     "            recon = None\n",
303 |     "        else:\n",
304 |     "            wave, sr = librosa.load(speaker_dicts[key][0], sr=24000)\n",
305 |     "            mel = preprocess(wave)\n",
306 |     "            c = mel.transpose(-1, -2).squeeze().to('cuda')\n",
307 |     "            recon = vocoder.inference(c)\n",
308 |     "            recon = recon.view(-1).cpu().numpy()\n",
309 |     "\n",
310 |     "    converted_samples[key] = y_out.numpy()\n",
311 |     "    reconstructed_samples[key] = recon\n",
312 |     "\n",
313 |     "    converted_mels[key] = out\n",
314 |     "    \n",
315 |     "    keys.append(key)\n",
316 |     "end = time.time()\n",
317 |     "print('total processing time: %.3f sec' % (end - start) )\n",
318 |     "\n",
319 |     "import IPython.display as ipd\n",
320 |     "for key, wave in converted_samples.items():\n",
321 |     "    print('Converted: %s' % key)\n",
322 |     "    display(ipd.Audio(wave, rate=24000))\n",
323 |     "    print('Reference (vocoder): %s' % key)\n",
324 |     "    if reconstructed_samples[key] is not None:\n",
325 |     "        display(ipd.Audio(reconstructed_samples[key], rate=24000))\n",
326 |     "\n",
327 |     "print('Original (vocoder):')\n",
328 |     "wave, sr = librosa.load(wav_path, sr=24000)\n",
329 |     "mel = preprocess(wave)\n",
330 |     "c = mel.transpose(-1, -2).squeeze().to('cuda')\n",
331 |     "with torch.no_grad():\n",
332 |     "    recon = vocoder.inference(c)\n",
333 |     "    recon = recon.view(-1).cpu().numpy()\n",
334 |     "display(ipd.Audio(recon, rate=24000))\n",
335 |     "print('Original:')\n",
336 |     "display(ipd.Audio(wav_path, rate=24000))"
337 |    ]
338 |   },
339 |   {
340 |    "cell_type": "markdown",
341 |    "metadata": {
342 |     "id": "SWh3o9hvGvJt"
343 |    },
344 |    "source": [
345 |     "#### Convert by mapping network"
346 |    ]
347 |   },
348 |   {
349 |    "cell_type": "code",
350 |    "execution_count": null,
351 |    "metadata": {},
352 |    "outputs": [],
353 |    "source": [
354 |     "# no reference, using mapping network\n",
355 |     "speaker_dicts = {}\n",
356 |     "selected_speakers = [273, 259, 258, 243, 254, 244, 236, 233, 230, 228]\n",
357 |     "for s in selected_speakers:\n",
358 |     "    k = s\n",
359 |     "    speaker_dicts['p' + str(s)] = ('', speakers.index(s))\n",
360 |     "\n",
361 |     "reference_embeddings = compute_style(speaker_dicts)"
362 |    ]
363 |   },
364 |   {
365 |    "cell_type": "code",
366 |    "execution_count": null,
367 |    "metadata": {
368 |     "scrolled": true
369 |    },
370 |    "outputs": [],
371 |    "source": [
372 |     "# conversion \n",
373 |     "import time\n",
374 |     "start = time.time()\n",
375 |     "    \n",
376 |     "source = preprocess(audio).to('cuda:0')\n",
377 |     "keys = []\n",
378 |     "converted_samples = {}\n",
379 |     "reconstructed_samples = {}\n",
380 |     "converted_mels = {}\n",
381 |     "\n",
382 |     "for key, (ref, _) in reference_embeddings.items():\n",
383 |     "    with torch.no_grad():\n",
384 |     "        f0_feat = F0_model.get_feature_GAN(source.unsqueeze(1))\n",
385 |     "        out = starganv2.generator(source.unsqueeze(1), ref, F0=f0_feat)\n",
386 |     "        \n",
387 |     "        c = out.transpose(-1, -2).squeeze().to('cuda')\n",
388 |     "        y_out = vocoder.inference(c)\n",
389 |     "        y_out = y_out.view(-1).cpu()\n",
390 |     "\n",
391 |     "        if key not in speaker_dicts or speaker_dicts[key][0] == \"\":\n",
392 |     "            recon = None\n",
393 |     "        else:\n",
394 |     "            wave, sr = librosa.load(speaker_dicts[key][0], sr=24000)\n",
395 |     "            mel = preprocess(wave)\n",
396 |     "            c = mel.transpose(-1, -2).squeeze().to('cuda')\n",
397 |     "            recon = vocoder.inference(c)\n",
398 |     "            recon = recon.view(-1).cpu().numpy()\n",
399 |     "\n",
400 |     "    converted_samples[key] = y_out.numpy()\n",
401 |     "    reconstructed_samples[key] = recon\n",
402 |     "\n",
403 |     "    converted_mels[key] = out\n",
404 |     "    \n",
405 |     "    keys.append(key)\n",
406 |     "end = time.time()\n",
407 |     "print('total processing time: %.3f sec' % (end - start) )\n",
408 |     "\n",
409 |     "import IPython.display as ipd\n",
410 |     "for key, wave in converted_samples.items():\n",
411 |     "    print('Converted: %s' % key)\n",
412 |     "    display(ipd.Audio(wave, rate=24000))\n",
413 |     "    print('Reference (vocoder): %s' % key)\n",
414 |     "    if reconstructed_samples[key] is not None:\n",
415 |     "        display(ipd.Audio(reconstructed_samples[key], rate=24000))\n",
416 |     "\n",
417 |     "print('Original (vocoder):')\n",
418 |     "wave, sr = librosa.load(wav_path, sr=24000)\n",
419 |     "mel = preprocess(wave)\n",
420 |     "c = mel.transpose(-1, -2).squeeze().to('cuda')\n",
421 |     "with torch.no_grad():\n",
422 |     "    recon = vocoder.inference(c)\n",
423 |     "    recon = recon.view(-1).cpu().numpy()\n",
424 |     "display(ipd.Audio(recon, rate=24000))\n",
425 |     "print('Original:')\n",
426 |     "display(ipd.Audio(wav_path, rate=24000))"
427 |    ]
428 |   }
429 |  ],
430 |  "metadata": {
431 |   "accelerator": "GPU",
432 |   "colab": {
433 |    "collapsed_sections": [
434 |     "hCpoXuZeGKAn"
435 |    ],
436 |    "name": "Starganv2_vc.ipynb",
437 |    "provenance": [],
438 |    "toc_visible": true
439 |   },
440 |   "kernelspec": {
441 |    "display_name": "Python 3",
442 |    "language": "python",
443 |    "name": "python3"
444 |   },
445 |   "language_info": {
446 |    "codemirror_mode": {
447 |     "name": "ipython",
448 |     "version": 3
449 |    },
450 |    "file_extension": ".py",
451 |    "mimetype": "text/x-python",
452 |    "name": "python",
453 |    "nbconvert_exporter": "python",
454 |    "pygments_lexer": "ipython3",
455 |    "version": "3.7.10"
456 |   }
457 |  },
458 |  "nbformat": 4,
459 |  "nbformat_minor": 1
460 | }
461 | 


--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
 1 | MIT License
 2 | 
 3 | Copyright (c) 2021 Aaron (Yinghao) Li
 4 | 
 5 | Permission is hereby granted, free of charge, to any person obtaining a copy
 6 | of this software and associated documentation files (the "Software"), to deal
 7 | in the Software without restriction, including without limitation the rights
 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 9 | copies of the Software, and to permit persons to whom the Software is
10 | furnished to do so, subject to the following conditions:
11 | 
12 | The above copyright notice and this permission notice shall be included in all
13 | copies or substantial portions of the Software.
14 | 
15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 | SOFTWARE.
22 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
 1 | # StarGANv2-VC: A Diverse, Unsupervised, Non-parallel Framework for Natural-Sounding Voice Conversion
 2 | 
 3 | ## For Hifi-GAN and BIGVGAN support, please check out [this branch](https://github.com/yl4579/StarGANv2-VC/tree/BIGVGAN). 
 4 | 
 5 | ### Yinghao Aaron Li, Ali Zare, Nima Mesgarani
 6 | 
 7 | > We present an unsupervised non-parallel many-to-many voice conversion (VC) method using a generative adversarial network (GAN) called StarGAN v2. Using a combination of adversarial source classifier loss and perceptual loss, our model significantly outperforms previous VC models. Although our model is trained only with 20 English speakers, it generalizes to a variety of voice conversion tasks, such as any-to-many, cross-lingual, and singing conversion. Using a style encoder, our framework can also convert plain reading speech into stylistic speech, such as emotional and falsetto speech. Subjective and objective evaluation experiments on a non-parallel many-to-many voice conversion task revealed that our model produces natural sounding voices, close to the sound quality of state-of-the-art text-tospeech (TTS) based voice conversion methods without the need for text labels. Moreover, our model is completely convolutional and with a faster-than-real-time vocoder such as Parallel WaveGAN can perform real-time voice conversion.
 8 | 
 9 | Paper: https://arxiv.org/abs/2107.10394
10 | 
11 | Audio samples: https://starganv2-vc.github.io/
12 | 
13 | ***Our work has been awarded [INTERSPEECH 2021 Best Paper Award](https://www.interspeech2021.org/best-student-paper-awards)**
14 | 
15 | ## Pre-requisites
16 | 1. Python >= 3.7
17 | 2. Clone this repository:
18 | ```bash
19 | git clone https://github.com/yl4579/StarGANv2-VC.git
20 | cd StarGANv2-VC
21 | ```
22 | 3. Install python requirements: 
23 | ```bash
24 | pip install SoundFile torchaudio munch parallel_wavegan torch pydub pyyaml click librosa
25 | ```
26 | 4. Download and extract the [VCTK dataset](https://datashare.ed.ac.uk/handle/10283/3443) 
27 | and use [VCTK.ipynb](https://github.com/yl4579/StarGANv2-VC/blob/main/Data/VCTK.ipynb) to prepare the data (downsample to 24 kHz etc.). You can also [download the dataset](https://drive.google.com/file/d/1t7QQbu4YC_P1mv9puA_KgSomSFDsSzD6/view?usp=sharing) we have prepared and unzip it to the `Data` folder, use the provided `config.yml` to reproduce our models. 
28 | 
29 | ## Training
30 | ```bash
31 | python train.py --config_path ./Configs/config.yml
32 | ```
33 | Please specify the training and validation data in `config.yml` file. Change `num_domains` to the number of speakers in the dataset. The data list format needs to be `filename.wav|speaker_number`, see [train_list.txt](https://github.com/yl4579/StarGANv2-VC/blob/main/Data/train_list.txt) as an example. 
34 | 
35 | Checkpoints and Tensorboard logs will be saved at `log_dir`. To speed up training, you may want to make `batch_size` as large as your GPU RAM can take. However, please note that `batch_size = 5` will take around 10G GPU RAM. 
36 | 
37 | ## Inference
38 | 
39 | Please refer to [inference.ipynb](https://github.com/yl4579/StarGANv2-VC/blob/main/Demo/inference.ipynb) for details. 
40 | 
41 | The pretrained StarGANv2 and ParallelWaveGAN on VCTK corpus can be downloaded at [StarGANv2 Link](https://huggingface.co/yl4579/StarGANv2-VC/blob/main/Models.zip) and [ParallelWaveGAN Link](https://huggingface.co/yl4579/StarGANv2-VC/blob/main/Vocoder.zip). Please unzip to `Models` and `Vocoder` respectivey and run each cell in the notebook.
42 | 
43 | ## ASR & F0 Models
44 | 
45 | The pretrained F0 and ASR models are provided under the `Utils` folder. Both the F0 and ASR models are trained with melspectrograms preprocessed using [meldataset.py](https://github.com/yl4579/StarGANv2-VC/blob/main/meldataset.py), and both models are trained on speech data only. 
46 | 
47 | The ASR model is trained on English corpus, but it appears to work when training StarGANv2 models in other languages such as Japanese. The F0 model also appears to work with singing data. For the best performance, however, training your own ASR and F0 models is encouraged for non-English and non-speech data. 
48 | 
49 | You can edit the [meldataset.py](https://github.com/yl4579/StarGANv2-VC/blob/main/meldataset.py) with your own melspectrogram preprocessing, but the provided pretrained models will no longer work. You will need to train your own ASR and F0 models with the new preprocessing. 
50 | 
51 | The code for training new ASR models is available [here](https://github.com/yl4579/AuxiliaryASR) and that for training new F0 models is available [here](https://github.com/yl4579/PitchExtractor).
52 | 
53 | ## References
54 | - [clovaai/stargan-v2](https://github.com/clovaai/stargan-v2)
55 | - [kan-bayashi/ParallelWaveGAN](https://github.com/kan-bayashi/ParallelWaveGAN)
56 | - [tosaka-m/japanese_realtime_tts](https://github.com/tosaka-m/japanese_realtime_tts)
57 | - [keums/melodyExtraction_JDC](https://github.com/keums/melodyExtraction_JDC)
58 | 
59 | ## Acknowledgement
60 | The author would like to thank [@tosaka-m](https://github.com/tosaka-m) for his great repository and valuable discussions.
61 | 


--------------------------------------------------------------------------------
/Utils/ASR/__init__.py:
--------------------------------------------------------------------------------
1 | 
2 | 


--------------------------------------------------------------------------------
/Utils/ASR/config.yml:
--------------------------------------------------------------------------------
 1 | log_dir: "logs"
 2 | save_freq: 20
 3 | device: "cuda"
 4 | epochs: 180
 5 | batch_size: 48
 6 | pretrained_model: ""
 7 | train_data: "asr_train_list.txt"
 8 | val_data: "asr_val_list.txt"
 9 | 
10 | dataset_params:
11 |   data_augmentation: true
12 | 
13 | preprocess_parasm:
14 |   sr: 24000
15 |   spect_params:
16 |     n_fft: 2048
17 |     win_length: 1200
18 |     hop_length: 300
19 |   mel_params:
20 |     n_mels: 80
21 | 
22 | model_params:
23 |    input_dim: 80
24 |    hidden_dim: 256
25 |    n_token: 80
26 |    token_embedding_dim: 256
27 | 
28 | optimizer_params:
29 |   lr: 0.0005


--------------------------------------------------------------------------------
/Utils/ASR/epoch_00100.pth:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yl4579/StarGANv2-VC/ebe97a0c41ea295dcea8f5e3e2b7e934edd01608/Utils/ASR/epoch_00100.pth


--------------------------------------------------------------------------------
/Utils/ASR/layers.py:
--------------------------------------------------------------------------------
  1 | import math
  2 | import torch
  3 | from torch import nn
  4 | from typing import Optional, Any
  5 | from torch import Tensor
  6 | import torch.nn.functional as F
  7 | import torchaudio
  8 | import torchaudio.functional as audio_F
  9 | 
 10 | import random
 11 | random.seed(0)
 12 | 
 13 | 
 14 | def _get_activation_fn(activ):
 15 |     if activ == 'relu':
 16 |         return nn.ReLU()
 17 |     elif activ == 'lrelu':
 18 |         return nn.LeakyReLU(0.2)
 19 |     elif activ == 'swish':
 20 |         return lambda x: x*torch.sigmoid(x)
 21 |     else:
 22 |         raise RuntimeError('Unexpected activ type %s, expected [relu, lrelu, swish]' % activ)
 23 | 
 24 | class LinearNorm(torch.nn.Module):
 25 |     def __init__(self, in_dim, out_dim, bias=True, w_init_gain='linear'):
 26 |         super(LinearNorm, self).__init__()
 27 |         self.linear_layer = torch.nn.Linear(in_dim, out_dim, bias=bias)
 28 | 
 29 |         torch.nn.init.xavier_uniform_(
 30 |             self.linear_layer.weight,
 31 |             gain=torch.nn.init.calculate_gain(w_init_gain))
 32 | 
 33 |     def forward(self, x):
 34 |         return self.linear_layer(x)
 35 | 
 36 | 
 37 | class ConvNorm(torch.nn.Module):
 38 |     def __init__(self, in_channels, out_channels, kernel_size=1, stride=1,
 39 |                  padding=None, dilation=1, bias=True, w_init_gain='linear', param=None):
 40 |         super(ConvNorm, self).__init__()
 41 |         if padding is None:
 42 |             assert(kernel_size % 2 == 1)
 43 |             padding = int(dilation * (kernel_size - 1) / 2)
 44 | 
 45 |         self.conv = torch.nn.Conv1d(in_channels, out_channels,
 46 |                                     kernel_size=kernel_size, stride=stride,
 47 |                                     padding=padding, dilation=dilation,
 48 |                                     bias=bias)
 49 | 
 50 |         torch.nn.init.xavier_uniform_(
 51 |             self.conv.weight, gain=torch.nn.init.calculate_gain(w_init_gain, param=param))
 52 | 
 53 |     def forward(self, signal):
 54 |         conv_signal = self.conv(signal)
 55 |         return conv_signal
 56 | 
 57 | class CausualConv(nn.Module):
 58 |     def __init__(self, in_channels, out_channels, kernel_size=1, stride=1, padding=1, dilation=1, bias=True, w_init_gain='linear', param=None):
 59 |         super(CausualConv, self).__init__()
 60 |         if padding is None:
 61 |             assert(kernel_size % 2 == 1)
 62 |             padding = int(dilation * (kernel_size - 1) / 2) * 2
 63 |         else:
 64 |             self.padding = padding * 2
 65 |         self.conv = nn.Conv1d(in_channels, out_channels,
 66 |                               kernel_size=kernel_size, stride=stride,
 67 |                               padding=self.padding,
 68 |                               dilation=dilation,
 69 |                               bias=bias)
 70 | 
 71 |         torch.nn.init.xavier_uniform_(
 72 |             self.conv.weight, gain=torch.nn.init.calculate_gain(w_init_gain, param=param))
 73 | 
 74 |     def forward(self, x):
 75 |         x = self.conv(x)
 76 |         x = x[:, :, :-self.padding]
 77 |         return x
 78 | 
 79 | class CausualBlock(nn.Module):
 80 |     def __init__(self, hidden_dim, n_conv=3, dropout_p=0.2, activ='lrelu'):
 81 |         super(CausualBlock, self).__init__()
 82 |         self.blocks = nn.ModuleList([
 83 |             self._get_conv(hidden_dim, dilation=3**i, activ=activ, dropout_p=dropout_p)
 84 |             for i in range(n_conv)])
 85 | 
 86 |     def forward(self, x):
 87 |         for block in self.blocks:
 88 |             res = x
 89 |             x = block(x)
 90 |             x += res
 91 |         return x
 92 | 
 93 |     def _get_conv(self, hidden_dim, dilation, activ='lrelu', dropout_p=0.2):
 94 |         layers = [
 95 |             CausualConv(hidden_dim, hidden_dim, kernel_size=3, padding=dilation, dilation=dilation),
 96 |             _get_activation_fn(activ),
 97 |             nn.BatchNorm1d(hidden_dim),
 98 |             nn.Dropout(p=dropout_p),
 99 |             CausualConv(hidden_dim, hidden_dim, kernel_size=3, padding=1, dilation=1),
100 |             _get_activation_fn(activ),
101 |             nn.Dropout(p=dropout_p)
102 |         ]
103 |         return nn.Sequential(*layers)
104 | 
105 | class ConvBlock(nn.Module):
106 |     def __init__(self, hidden_dim, n_conv=3, dropout_p=0.2, activ='relu'):
107 |         super().__init__()
108 |         self._n_groups = 8
109 |         self.blocks = nn.ModuleList([
110 |             self._get_conv(hidden_dim, dilation=3**i, activ=activ, dropout_p=dropout_p)
111 |             for i in range(n_conv)])
112 | 
113 | 
114 |     def forward(self, x):
115 |         for block in self.blocks:
116 |             res = x
117 |             x = block(x)
118 |             x += res
119 |         return x
120 | 
121 |     def _get_conv(self, hidden_dim, dilation, activ='relu', dropout_p=0.2):
122 |         layers = [
123 |             ConvNorm(hidden_dim, hidden_dim, kernel_size=3, padding=dilation, dilation=dilation),
124 |             _get_activation_fn(activ),
125 |             nn.GroupNorm(num_groups=self._n_groups, num_channels=hidden_dim),
126 |             nn.Dropout(p=dropout_p),
127 |             ConvNorm(hidden_dim, hidden_dim, kernel_size=3, padding=1, dilation=1),
128 |             _get_activation_fn(activ),
129 |             nn.Dropout(p=dropout_p)
130 |         ]
131 |         return nn.Sequential(*layers)
132 | 
133 | class LocationLayer(nn.Module):
134 |     def __init__(self, attention_n_filters, attention_kernel_size,
135 |                  attention_dim):
136 |         super(LocationLayer, self).__init__()
137 |         padding = int((attention_kernel_size - 1) / 2)
138 |         self.location_conv = ConvNorm(2, attention_n_filters,
139 |                                       kernel_size=attention_kernel_size,
140 |                                       padding=padding, bias=False, stride=1,
141 |                                       dilation=1)
142 |         self.location_dense = LinearNorm(attention_n_filters, attention_dim,
143 |                                          bias=False, w_init_gain='tanh')
144 | 
145 |     def forward(self, attention_weights_cat):
146 |         processed_attention = self.location_conv(attention_weights_cat)
147 |         processed_attention = processed_attention.transpose(1, 2)
148 |         processed_attention = self.location_dense(processed_attention)
149 |         return processed_attention
150 | 
151 | 
152 | class Attention(nn.Module):
153 |     def __init__(self, attention_rnn_dim, embedding_dim, attention_dim,
154 |                  attention_location_n_filters, attention_location_kernel_size):
155 |         super(Attention, self).__init__()
156 |         self.query_layer = LinearNorm(attention_rnn_dim, attention_dim,
157 |                                       bias=False, w_init_gain='tanh')
158 |         self.memory_layer = LinearNorm(embedding_dim, attention_dim, bias=False,
159 |                                        w_init_gain='tanh')
160 |         self.v = LinearNorm(attention_dim, 1, bias=False)
161 |         self.location_layer = LocationLayer(attention_location_n_filters,
162 |                                             attention_location_kernel_size,
163 |                                             attention_dim)
164 |         self.score_mask_value = -float("inf")
165 | 
166 |     def get_alignment_energies(self, query, processed_memory,
167 |                                attention_weights_cat):
168 |         """
169 |         PARAMS
170 |         ------
171 |         query: decoder output (batch, n_mel_channels * n_frames_per_step)
172 |         processed_memory: processed encoder outputs (B, T_in, attention_dim)
173 |         attention_weights_cat: cumulative and prev. att weights (B, 2, max_time)
174 |         RETURNS
175 |         -------
176 |         alignment (batch, max_time)
177 |         """
178 | 
179 |         processed_query = self.query_layer(query.unsqueeze(1))
180 |         processed_attention_weights = self.location_layer(attention_weights_cat)
181 |         energies = self.v(torch.tanh(
182 |             processed_query + processed_attention_weights + processed_memory))
183 | 
184 |         energies = energies.squeeze(-1)
185 |         return energies
186 | 
187 |     def forward(self, attention_hidden_state, memory, processed_memory,
188 |                 attention_weights_cat, mask):
189 |         """
190 |         PARAMS
191 |         ------
192 |         attention_hidden_state: attention rnn last output
193 |         memory: encoder outputs
194 |         processed_memory: processed encoder outputs
195 |         attention_weights_cat: previous and cummulative attention weights
196 |         mask: binary mask for padded data
197 |         """
198 |         alignment = self.get_alignment_energies(
199 |             attention_hidden_state, processed_memory, attention_weights_cat)
200 | 
201 |         if mask is not None:
202 |             alignment.data.masked_fill_(mask, self.score_mask_value)
203 | 
204 |         attention_weights = F.softmax(alignment, dim=1)
205 |         attention_context = torch.bmm(attention_weights.unsqueeze(1), memory)
206 |         attention_context = attention_context.squeeze(1)
207 | 
208 |         return attention_context, attention_weights
209 | 
210 | 
211 | class ForwardAttentionV2(nn.Module):
212 |     def __init__(self, attention_rnn_dim, embedding_dim, attention_dim,
213 |                  attention_location_n_filters, attention_location_kernel_size):
214 |         super(ForwardAttentionV2, self).__init__()
215 |         self.query_layer = LinearNorm(attention_rnn_dim, attention_dim,
216 |                                       bias=False, w_init_gain='tanh')
217 |         self.memory_layer = LinearNorm(embedding_dim, attention_dim, bias=False,
218 |                                        w_init_gain='tanh')
219 |         self.v = LinearNorm(attention_dim, 1, bias=False)
220 |         self.location_layer = LocationLayer(attention_location_n_filters,
221 |                                             attention_location_kernel_size,
222 |                                             attention_dim)
223 |         self.score_mask_value = -float(1e20)
224 | 
225 |     def get_alignment_energies(self, query, processed_memory,
226 |                                attention_weights_cat):
227 |         """
228 |         PARAMS
229 |         ------
230 |         query: decoder output (batch, n_mel_channels * n_frames_per_step)
231 |         processed_memory: processed encoder outputs (B, T_in, attention_dim)
232 |         attention_weights_cat:  prev. and cumulative att weights (B, 2, max_time)
233 |         RETURNS
234 |         -------
235 |         alignment (batch, max_time)
236 |         """
237 | 
238 |         processed_query = self.query_layer(query.unsqueeze(1))
239 |         processed_attention_weights = self.location_layer(attention_weights_cat)
240 |         energies = self.v(torch.tanh(
241 |             processed_query + processed_attention_weights + processed_memory))
242 | 
243 |         energies = energies.squeeze(-1)
244 |         return energies
245 | 
246 |     def forward(self, attention_hidden_state, memory, processed_memory,
247 |                 attention_weights_cat, mask, log_alpha):
248 |         """
249 |         PARAMS
250 |         ------
251 |         attention_hidden_state: attention rnn last output
252 |         memory: encoder outputs
253 |         processed_memory: processed encoder outputs
254 |         attention_weights_cat: previous and cummulative attention weights
255 |         mask: binary mask for padded data
256 |         """
257 |         log_energy = self.get_alignment_energies(
258 |             attention_hidden_state, processed_memory, attention_weights_cat)
259 | 
260 |         #log_energy =
261 | 
262 |         if mask is not None:
263 |             log_energy.data.masked_fill_(mask, self.score_mask_value)
264 | 
265 |         #attention_weights = F.softmax(alignment, dim=1)
266 | 
267 |         #content_score = log_energy.unsqueeze(1) #[B, MAX_TIME] -> [B, 1, MAX_TIME]
268 |         #log_alpha = log_alpha.unsqueeze(2) #[B, MAX_TIME] -> [B, MAX_TIME, 1]
269 | 
270 |         #log_total_score = log_alpha + content_score
271 | 
272 |         #previous_attention_weights = attention_weights_cat[:,0,:]
273 | 
274 |         log_alpha_shift_padded = []
275 |         max_time = log_energy.size(1)
276 |         for sft in range(2):
277 |             shifted = log_alpha[:,:max_time-sft]
278 |             shift_padded = F.pad(shifted, (sft,0), 'constant', self.score_mask_value)
279 |             log_alpha_shift_padded.append(shift_padded.unsqueeze(2))
280 | 
281 |         biased = torch.logsumexp(torch.cat(log_alpha_shift_padded,2), 2)
282 | 
283 |         log_alpha_new = biased +  log_energy
284 | 
285 |         attention_weights =  F.softmax(log_alpha_new, dim=1)
286 | 
287 |         attention_context = torch.bmm(attention_weights.unsqueeze(1), memory)
288 |         attention_context = attention_context.squeeze(1)
289 | 
290 |         return attention_context, attention_weights, log_alpha_new
291 | 
292 | 
293 | class PhaseShuffle2d(nn.Module):
294 |     def __init__(self, n=2):
295 |         super(PhaseShuffle2d, self).__init__()
296 |         self.n = n
297 |         self.random = random.Random(1)
298 | 
299 |     def forward(self, x, move=None):
300 |         # x.size = (B, C, M, L)
301 |         if move is None:
302 |             move = self.random.randint(-self.n, self.n)
303 | 
304 |         if move == 0:
305 |             return x
306 |         else:
307 |             left = x[:, :, :, :move]
308 |             right = x[:, :, :, move:]
309 |             shuffled = torch.cat([right, left], dim=3)
310 |         return shuffled
311 | 
312 | class PhaseShuffle1d(nn.Module):
313 |     def __init__(self, n=2):
314 |         super(PhaseShuffle1d, self).__init__()
315 |         self.n = n
316 |         self.random = random.Random(1)
317 | 
318 |     def forward(self, x, move=None):
319 |         # x.size = (B, C, M, L)
320 |         if move is None:
321 |             move = self.random.randint(-self.n, self.n)
322 | 
323 |         if move == 0:
324 |             return x
325 |         else:
326 |             left = x[:, :,  :move]
327 |             right = x[:, :, move:]
328 |             shuffled = torch.cat([right, left], dim=2)
329 | 
330 |         return shuffled
331 | 
332 | class MFCC(nn.Module):
333 |     def __init__(self, n_mfcc=40, n_mels=80):
334 |         super(MFCC, self).__init__()
335 |         self.n_mfcc = n_mfcc
336 |         self.n_mels = n_mels
337 |         self.norm = 'ortho'
338 |         dct_mat = audio_F.create_dct(self.n_mfcc, self.n_mels, self.norm)
339 |         self.register_buffer('dct_mat', dct_mat)
340 | 
341 |     def forward(self, mel_specgram):
342 |         if len(mel_specgram.shape) == 2:
343 |             mel_specgram = mel_specgram.unsqueeze(0)
344 |             unsqueezed = True
345 |         else:
346 |             unsqueezed = False
347 |         # (channel, n_mels, time).tranpose(...) dot (n_mels, n_mfcc)
348 |         # -> (channel, time, n_mfcc).tranpose(...)
349 |         mfcc = torch.matmul(mel_specgram.transpose(1, 2), self.dct_mat).transpose(1, 2)
350 | 
351 |         # unpack batch
352 |         if unsqueezed:
353 |             mfcc = mfcc.squeeze(0)
354 |         return mfcc
355 | 


--------------------------------------------------------------------------------
/Utils/ASR/models.py:
--------------------------------------------------------------------------------
  1 | import math
  2 | import torch
  3 | from torch import nn
  4 | from torch.nn import TransformerEncoder
  5 | import torch.nn.functional as F
  6 | from .layers import MFCC, Attention, LinearNorm, ConvNorm, ConvBlock
  7 | 
  8 | class ASRCNN(nn.Module):
  9 |     def __init__(self,
 10 |                  input_dim=80,
 11 |                  hidden_dim=256,
 12 |                  n_token=35,
 13 |                  n_layers=6,
 14 |                  token_embedding_dim=256,
 15 | 
 16 |     ):
 17 |         super().__init__()
 18 |         self.n_token = n_token
 19 |         self.n_down = 1
 20 |         self.to_mfcc = MFCC()
 21 |         self.init_cnn = ConvNorm(input_dim//2, hidden_dim, kernel_size=7, padding=3, stride=2)
 22 |         self.cnns = nn.Sequential(
 23 |             *[nn.Sequential(
 24 |                 ConvBlock(hidden_dim),
 25 |                 nn.GroupNorm(num_groups=1, num_channels=hidden_dim)
 26 |             ) for n in range(n_layers)])
 27 |         self.projection = ConvNorm(hidden_dim, hidden_dim // 2)
 28 |         self.ctc_linear = nn.Sequential(
 29 |             LinearNorm(hidden_dim//2, hidden_dim),
 30 |             nn.ReLU(),
 31 |             LinearNorm(hidden_dim, n_token))
 32 |         self.asr_s2s = ASRS2S(
 33 |             embedding_dim=token_embedding_dim,
 34 |             hidden_dim=hidden_dim//2,
 35 |             n_token=n_token)
 36 | 
 37 |     def forward(self, x, src_key_padding_mask=None, text_input=None):
 38 |         x = self.to_mfcc(x)
 39 |         x = self.init_cnn(x)
 40 |         x = self.cnns(x)
 41 |         x = self.projection(x)
 42 |         x = x.transpose(1, 2)
 43 |         ctc_logit = self.ctc_linear(x)
 44 |         if text_input is not None:
 45 |             _, s2s_logit, s2s_attn = self.asr_s2s(x, src_key_padding_mask, text_input)
 46 |             return ctc_logit, s2s_logit, s2s_attn
 47 |         else:
 48 |             return ctc_logit
 49 | 
 50 |     def get_feature(self, x):
 51 |         x = self.to_mfcc(x.squeeze(1))
 52 |         x = self.init_cnn(x)
 53 |         x = self.cnns(x)
 54 |         x = self.projection(x)
 55 |         return x
 56 | 
 57 |     def length_to_mask(self, lengths):
 58 |         mask = torch.arange(lengths.max()).unsqueeze(0).expand(lengths.shape[0], -1).type_as(lengths)
 59 |         mask = torch.gt(mask+1, lengths.unsqueeze(1)).to(lengths.device)
 60 |         return mask
 61 | 
 62 |     def get_future_mask(self, out_length, unmask_future_steps=0):
 63 |         """
 64 |         Args:
 65 |             out_length (int): returned mask shape is (out_length, out_length).
 66 |             unmask_futre_steps (int): unmasking future step size.
 67 |         Return:
 68 |             mask (torch.BoolTensor): mask future timesteps mask[i, j] = True if i > j + unmask_future_steps else False
 69 |         """
 70 |         index_tensor = torch.arange(out_length).unsqueeze(0).expand(out_length, -1)
 71 |         mask = torch.gt(index_tensor, index_tensor.T + unmask_future_steps)
 72 |         return mask
 73 | 
 74 | class ASRS2S(nn.Module):
 75 |     def __init__(self,
 76 |                  embedding_dim=256,
 77 |                  hidden_dim=512,
 78 |                  n_location_filters=32,
 79 |                  location_kernel_size=63,
 80 |                  n_token=40):
 81 |         super(ASRS2S, self).__init__()
 82 |         self.embedding = nn.Embedding(n_token, embedding_dim)
 83 |         val_range = math.sqrt(6 / hidden_dim)
 84 |         self.embedding.weight.data.uniform_(-val_range, val_range)
 85 | 
 86 |         self.decoder_rnn_dim = hidden_dim
 87 |         self.project_to_n_symbols = nn.Linear(self.decoder_rnn_dim, n_token)
 88 |         self.attention_layer = Attention(
 89 |             self.decoder_rnn_dim,
 90 |             hidden_dim,
 91 |             hidden_dim,
 92 |             n_location_filters,
 93 |             location_kernel_size
 94 |         )
 95 |         self.decoder_rnn = nn.LSTMCell(self.decoder_rnn_dim + embedding_dim, self.decoder_rnn_dim)
 96 |         self.project_to_hidden = nn.Sequential(
 97 |             LinearNorm(self.decoder_rnn_dim * 2, hidden_dim),
 98 |             nn.Tanh())
 99 |         self.sos = 1
100 |         self.eos = 2
101 | 
102 |     def initialize_decoder_states(self, memory, mask):
103 |         """
104 |         moemory.shape = (B, L, H) = (Batchsize, Maxtimestep, Hiddendim)
105 |         """
106 |         B, L, H = memory.shape
107 |         self.decoder_hidden = torch.zeros((B, self.decoder_rnn_dim)).type_as(memory)
108 |         self.decoder_cell = torch.zeros((B, self.decoder_rnn_dim)).type_as(memory)
109 |         self.attention_weights = torch.zeros((B, L)).type_as(memory)
110 |         self.attention_weights_cum = torch.zeros((B, L)).type_as(memory)
111 |         self.attention_context = torch.zeros((B, H)).type_as(memory)
112 |         self.memory = memory
113 |         self.processed_memory = self.attention_layer.memory_layer(memory)
114 |         self.mask = mask
115 |         self.unk_index = 3
116 |         self.random_mask = 0.1
117 | 
118 |     def forward(self, memory, memory_mask, text_input):
119 |         """
120 |         moemory.shape = (B, L, H) = (Batchsize, Maxtimestep, Hiddendim)
121 |         moemory_mask.shape = (B, L, )
122 |         texts_input.shape = (B, T)
123 |         """
124 |         self.initialize_decoder_states(memory, memory_mask)
125 |         # text random mask
126 |         random_mask = (torch.rand(text_input.shape) < self.random_mask).to(text_input.device)
127 |         _text_input = text_input.clone()
128 |         _text_input.masked_fill_(random_mask, self.unk_index)
129 |         decoder_inputs = self.embedding(_text_input).transpose(0, 1) # -> [T, B, channel]
130 |         start_embedding = self.embedding(
131 |             torch.LongTensor([self.sos]*decoder_inputs.size(1)).to(decoder_inputs.device))
132 |         decoder_inputs = torch.cat((start_embedding.unsqueeze(0), decoder_inputs), dim=0)
133 | 
134 |         hidden_outputs, logit_outputs, alignments = [], [], []
135 |         while len(hidden_outputs) < decoder_inputs.size(0):
136 | 
137 |             decoder_input = decoder_inputs[len(hidden_outputs)]
138 |             hidden, logit, attention_weights = self.decode(decoder_input)
139 |             hidden_outputs += [hidden]
140 |             logit_outputs += [logit]
141 |             alignments += [attention_weights]
142 | 
143 |         hidden_outputs, logit_outputs, alignments = \
144 |             self.parse_decoder_outputs(
145 |                 hidden_outputs, logit_outputs, alignments)
146 | 
147 |         return hidden_outputs, logit_outputs, alignments
148 | 
149 | 
150 |     def decode(self, decoder_input):
151 | 
152 |         cell_input = torch.cat((decoder_input, self.attention_context), -1)
153 |         self.decoder_hidden, self.decoder_cell = self.decoder_rnn(
154 |             cell_input,
155 |             (self.decoder_hidden, self.decoder_cell))
156 | 
157 |         attention_weights_cat = torch.cat(
158 |             (self.attention_weights.unsqueeze(1),
159 |             self.attention_weights_cum.unsqueeze(1)),dim=1)
160 | 
161 |         self.attention_context, self.attention_weights = self.attention_layer(
162 |             self.decoder_hidden,
163 |             self.memory,
164 |             self.processed_memory,
165 |             attention_weights_cat,
166 |             self.mask)
167 | 
168 |         self.attention_weights_cum += self.attention_weights
169 | 
170 |         hidden_and_context = torch.cat((self.decoder_hidden, self.attention_context), -1)
171 |         hidden = self.project_to_hidden(hidden_and_context)
172 | 
173 |         # dropout to increasing g
174 |         logit = self.project_to_n_symbols(F.dropout(hidden, 0.5, self.training))
175 | 
176 |         return hidden, logit, self.attention_weights
177 | 
178 |     def parse_decoder_outputs(self, hidden, logit, alignments):
179 | 
180 |         # -> [B, T_out + 1, max_time]
181 |         alignments = torch.stack(alignments).transpose(0,1)
182 |         # [T_out + 1, B, n_symbols] -> [B, T_out + 1,  n_symbols]
183 |         logit = torch.stack(logit).transpose(0, 1).contiguous()
184 |         hidden = torch.stack(hidden).transpose(0, 1).contiguous()
185 | 
186 |         return hidden, logit, alignments
187 | 


--------------------------------------------------------------------------------
/Utils/JDC/__init__.py:
--------------------------------------------------------------------------------
1 | 
2 | 


--------------------------------------------------------------------------------
/Utils/JDC/bst.t7:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/yl4579/StarGANv2-VC/ebe97a0c41ea295dcea8f5e3e2b7e934edd01608/Utils/JDC/bst.t7


--------------------------------------------------------------------------------
/Utils/JDC/model.py:
--------------------------------------------------------------------------------
  1 | """
  2 | Implementation of model from:
  3 | Kum et al. - "Joint Detection and Classification of Singing Voice Melody Using
  4 | Convolutional Recurrent Neural Networks" (2019)
  5 | Link: https://www.semanticscholar.org/paper/Joint-Detection-and-Classification-of-Singing-Voice-Kum-Nam/60a2ad4c7db43bace75805054603747fcd062c0d
  6 | """
  7 | import torch
  8 | from torch import nn
  9 |         
 10 | class JDCNet(nn.Module):
 11 |     """
 12 |     Joint Detection and Classification Network model for singing voice melody.
 13 |     """
 14 |     def __init__(self, num_class=722, seq_len=31, leaky_relu_slope=0.01):
 15 |         super().__init__()
 16 |         self.seq_len = seq_len  # 31
 17 |         self.num_class = num_class
 18 | 
 19 |         # input = (b, 1, 31, 513), b = batch size
 20 |         self.conv_block = nn.Sequential(
 21 |             nn.Conv2d(in_channels=1, out_channels=64, kernel_size=3, padding=1, bias=False),  # out: (b, 64, 31, 513)
 22 |             nn.BatchNorm2d(num_features=64),
 23 |             nn.LeakyReLU(leaky_relu_slope, inplace=True),
 24 |             nn.Conv2d(64, 64, 3, padding=1, bias=False),  # (b, 64, 31, 513)
 25 |         )
 26 | 
 27 |         # res blocks
 28 |         self.res_block1 = ResBlock(in_channels=64, out_channels=128)  # (b, 128, 31, 128)
 29 |         self.res_block2 = ResBlock(in_channels=128, out_channels=192)  # (b, 192, 31, 32)
 30 |         self.res_block3 = ResBlock(in_channels=192, out_channels=256)  # (b, 256, 31, 8)
 31 | 
 32 |         # pool block
 33 |         self.pool_block = nn.Sequential(
 34 |             nn.BatchNorm2d(num_features=256),
 35 |             nn.LeakyReLU(leaky_relu_slope, inplace=True),
 36 |             nn.MaxPool2d(kernel_size=(1, 4)),  # (b, 256, 31, 2)
 37 |             nn.Dropout(p=0.5),
 38 |         )
 39 | 
 40 |         # maxpool layers (for auxiliary network inputs)
 41 |         # in = (b, 128, 31, 513) from conv_block, out = (b, 128, 31, 2)
 42 |         self.maxpool1 = nn.MaxPool2d(kernel_size=(1, 40))
 43 |         # in = (b, 128, 31, 128) from res_block1, out = (b, 128, 31, 2)
 44 |         self.maxpool2 = nn.MaxPool2d(kernel_size=(1, 20))
 45 |         # in = (b, 128, 31, 32) from res_block2, out = (b, 128, 31, 2)
 46 |         self.maxpool3 = nn.MaxPool2d(kernel_size=(1, 10))
 47 | 
 48 |         # in = (b, 640, 31, 2), out = (b, 256, 31, 2)
 49 |         self.detector_conv = nn.Sequential(
 50 |             nn.Conv2d(640, 256, 1, bias=False),
 51 |             nn.BatchNorm2d(256),
 52 |             nn.LeakyReLU(leaky_relu_slope, inplace=True),
 53 |             nn.Dropout(p=0.5),
 54 |         )
 55 | 
 56 |         # input: (b, 31, 512) - resized from (b, 256, 31, 2)
 57 |         self.bilstm_classifier = nn.LSTM(
 58 |             input_size=512, hidden_size=256,
 59 |             batch_first=True, bidirectional=True)  # (b, 31, 512)
 60 | 
 61 |         # input: (b, 31, 512) - resized from (b, 256, 31, 2)
 62 |         self.bilstm_detector = nn.LSTM(
 63 |             input_size=512, hidden_size=256,
 64 |             batch_first=True, bidirectional=True)  # (b, 31, 512)
 65 | 
 66 |         # input: (b * 31, 512)
 67 |         self.classifier = nn.Linear(in_features=512, out_features=self.num_class)  # (b * 31, num_class)
 68 | 
 69 |         # input: (b * 31, 512)
 70 |         self.detector = nn.Linear(in_features=512, out_features=2)  # (b * 31, 2) - binary classifier
 71 | 
 72 |         # initialize weights
 73 |         self.apply(self.init_weights)
 74 | 
 75 |     def get_feature_GAN(self, x):
 76 |         seq_len = x.shape[-2]
 77 |         x = x.float().transpose(-1, -2)
 78 |         
 79 |         convblock_out = self.conv_block(x)
 80 |         
 81 |         resblock1_out = self.res_block1(convblock_out)
 82 |         resblock2_out = self.res_block2(resblock1_out)
 83 |         resblock3_out = self.res_block3(resblock2_out)
 84 |         poolblock_out = self.pool_block[0](resblock3_out)
 85 |         poolblock_out = self.pool_block[1](poolblock_out)
 86 |         
 87 |         return poolblock_out.transpose(-1, -2)
 88 |         
 89 |     def forward(self, x):
 90 |         """
 91 |         Returns:
 92 |             classification_prediction, detection_prediction
 93 |             sizes: (b, 31, 722), (b, 31, 2)
 94 |         """
 95 |         ###############################
 96 |         # forward pass for classifier #
 97 |         ###############################
 98 |         x = x.float().transpose(-1, -2)
 99 |         
100 |         convblock_out = self.conv_block(x)
101 |         
102 |         resblock1_out = self.res_block1(convblock_out)
103 |         resblock2_out = self.res_block2(resblock1_out)
104 |         resblock3_out = self.res_block3(resblock2_out)
105 |         
106 |         
107 |         poolblock_out = self.pool_block[0](resblock3_out)
108 |         poolblock_out = self.pool_block[1](poolblock_out)
109 |         GAN_feature = poolblock_out.transpose(-1, -2)
110 |         poolblock_out = self.pool_block[2](poolblock_out)
111 |         
112 |         # (b, 256, 31, 2) => (b, 31, 256, 2) => (b, 31, 512)
113 |         classifier_out = poolblock_out.permute(0, 2, 1, 3).contiguous().view((-1, self.seq_len, 512))
114 |         classifier_out, _ = self.bilstm_classifier(classifier_out)  # ignore the hidden states
115 | 
116 |         classifier_out = classifier_out.contiguous().view((-1, 512))  # (b * 31, 512)
117 |         classifier_out = self.classifier(classifier_out)
118 |         classifier_out = classifier_out.view((-1, self.seq_len, self.num_class))  # (b, 31, num_class)
119 |         
120 |         # sizes: (b, 31, 722), (b, 31, 2)
121 |         # classifier output consists of predicted pitch classes per frame
122 |         # detector output consists of: (isvoice, notvoice) estimates per frame
123 |         return torch.abs(classifier_out.squeeze()), GAN_feature, poolblock_out
124 | 
125 |     @staticmethod
126 |     def init_weights(m):
127 |         if isinstance(m, nn.Linear):
128 |             nn.init.kaiming_uniform_(m.weight)
129 |             if m.bias is not None:
130 |                 nn.init.constant_(m.bias, 0)
131 |         elif isinstance(m, nn.Conv2d):
132 |             nn.init.xavier_normal_(m.weight)
133 |         elif isinstance(m, nn.LSTM) or isinstance(m, nn.LSTMCell):
134 |             for p in m.parameters():
135 |                 if p.data is None:
136 |                     continue
137 | 
138 |                 if len(p.shape) >= 2:
139 |                     nn.init.orthogonal_(p.data)
140 |                 else:
141 |                     nn.init.normal_(p.data)
142 |                     
143 | 
144 | class ResBlock(nn.Module):
145 |     def __init__(self, in_channels: int, out_channels: int, leaky_relu_slope=0.01):
146 |         super().__init__()
147 |         self.downsample = in_channels != out_channels
148 | 
149 |         # BN / LReLU / MaxPool layer before the conv layer - see Figure 1b in the paper
150 |         self.pre_conv = nn.Sequential(
151 |             nn.BatchNorm2d(num_features=in_channels),
152 |             nn.LeakyReLU(leaky_relu_slope, inplace=True),
153 |             nn.MaxPool2d(kernel_size=(1, 2)),  # apply downsampling on the y axis only
154 |         )
155 | 
156 |         # conv layers
157 |         self.conv = nn.Sequential(
158 |             nn.Conv2d(in_channels=in_channels, out_channels=out_channels,
159 |                       kernel_size=3, padding=1, bias=False),
160 |             nn.BatchNorm2d(out_channels),
161 |             nn.LeakyReLU(leaky_relu_slope, inplace=True),
162 |             nn.Conv2d(out_channels, out_channels, 3, padding=1, bias=False),
163 |         )
164 | 
165 |         # 1 x 1 convolution layer to match the feature dimensions
166 |         self.conv1by1 = None
167 |         if self.downsample:
168 |             self.conv1by1 = nn.Conv2d(in_channels, out_channels, 1, bias=False)
169 | 
170 |     def forward(self, x):
171 |         x = self.pre_conv(x)
172 |         if self.downsample:
173 |             x = self.conv(x) + self.conv1by1(x)
174 |         else:
175 |             x = self.conv(x) + x
176 |         return x


--------------------------------------------------------------------------------
/Utils/__init__.py:
--------------------------------------------------------------------------------
1 | 
2 | 


--------------------------------------------------------------------------------
/losses.py:
--------------------------------------------------------------------------------
  1 | #coding:utf-8
  2 | 
  3 | import os
  4 | import torch
  5 | 
  6 | from torch import nn
  7 | from munch import Munch
  8 | from transforms import build_transforms
  9 | 
 10 | import torch.nn.functional as F
 11 | import numpy as np
 12 | 
 13 | def compute_d_loss(nets, args, x_real, y_org, y_trg, z_trg=None, x_ref=None, use_r1_reg=True, use_adv_cls=False, use_con_reg=False):
 14 |     args = Munch(args)
 15 | 
 16 |     assert (z_trg is None) != (x_ref is None)
 17 |     # with real audios
 18 |     x_real.requires_grad_()
 19 |     out = nets.discriminator(x_real, y_org)
 20 |     loss_real = adv_loss(out, 1)
 21 |     
 22 |     # R1 regularizaition (https://arxiv.org/abs/1801.04406v4)
 23 |     if use_r1_reg:
 24 |         loss_reg = r1_reg(out, x_real)
 25 |     else:
 26 |         loss_reg = torch.FloatTensor([0]).to(x_real.device)
 27 |     
 28 |     # consistency regularization (bCR-GAN: https://arxiv.org/abs/2002.04724)
 29 |     loss_con_reg = torch.FloatTensor([0]).to(x_real.device)
 30 |     if use_con_reg:
 31 |         t = build_transforms()
 32 |         out_aug = nets.discriminator(t(x_real).detach(), y_org)
 33 |         loss_con_reg += F.smooth_l1_loss(out, out_aug)
 34 |     
 35 |     # with fake audios
 36 |     with torch.no_grad():
 37 |         if z_trg is not None:
 38 |             s_trg = nets.mapping_network(z_trg, y_trg)
 39 |         else:  # x_ref is not None
 40 |             s_trg = nets.style_encoder(x_ref, y_trg)
 41 |             
 42 |         F0 = nets.f0_model.get_feature_GAN(x_real)
 43 |         x_fake = nets.generator(x_real, s_trg, masks=None, F0=F0)
 44 |     out = nets.discriminator(x_fake, y_trg)
 45 |     loss_fake = adv_loss(out, 0)
 46 |     if use_con_reg:
 47 |         out_aug = nets.discriminator(t(x_fake).detach(), y_trg)
 48 |         loss_con_reg += F.smooth_l1_loss(out, out_aug)
 49 |     
 50 |     # adversarial classifier loss
 51 |     if use_adv_cls:
 52 |         out_de = nets.discriminator.classifier(x_fake)
 53 |         loss_real_adv_cls = F.cross_entropy(out_de[y_org != y_trg], y_org[y_org != y_trg])
 54 |         
 55 |         if use_con_reg:
 56 |             out_de_aug = nets.discriminator.classifier(t(x_fake).detach())
 57 |             loss_con_reg += F.smooth_l1_loss(out_de, out_de_aug)
 58 |     else:
 59 |         loss_real_adv_cls = torch.zeros(1).mean()
 60 |         
 61 |     loss = loss_real + loss_fake + args.lambda_reg * loss_reg + \
 62 |             args.lambda_adv_cls * loss_real_adv_cls + \
 63 |             args.lambda_con_reg * loss_con_reg 
 64 | 
 65 |     return loss, Munch(real=loss_real.item(),
 66 |                        fake=loss_fake.item(),
 67 |                        reg=loss_reg.item(),
 68 |                        real_adv_cls=loss_real_adv_cls.item(),
 69 |                        con_reg=loss_con_reg.item())
 70 | 
 71 | def compute_g_loss(nets, args, x_real, y_org, y_trg, z_trgs=None, x_refs=None, use_adv_cls=False):
 72 |     args = Munch(args)
 73 |     
 74 |     assert (z_trgs is None) != (x_refs is None)
 75 |     if z_trgs is not None:
 76 |         z_trg, z_trg2 = z_trgs
 77 |     if x_refs is not None:
 78 |         x_ref, x_ref2 = x_refs
 79 |         
 80 |     # compute style vectors
 81 |     if z_trgs is not None:
 82 |         s_trg = nets.mapping_network(z_trg, y_trg)
 83 |     else:
 84 |         s_trg = nets.style_encoder(x_ref, y_trg)
 85 |     
 86 |     # compute ASR/F0 features (real)
 87 |     with torch.no_grad():
 88 |         F0_real, GAN_F0_real, cyc_F0_real = nets.f0_model(x_real)
 89 |         ASR_real = nets.asr_model.get_feature(x_real)
 90 |     
 91 |     # adversarial loss
 92 |     x_fake = nets.generator(x_real, s_trg, masks=None, F0=GAN_F0_real)
 93 |     out = nets.discriminator(x_fake, y_trg) 
 94 |     loss_adv = adv_loss(out, 1)
 95 |     
 96 |     # compute ASR/F0 features (fake)
 97 |     F0_fake, GAN_F0_fake, _ = nets.f0_model(x_fake)
 98 |     ASR_fake = nets.asr_model.get_feature(x_fake)
 99 |     
100 |     # norm consistency loss
101 |     x_fake_norm = log_norm(x_fake)
102 |     x_real_norm = log_norm(x_real)
103 |     loss_norm = ((torch.nn.ReLU()(torch.abs(x_fake_norm - x_real_norm) - args.norm_bias))**2).mean()
104 |     
105 |     # F0 loss
106 |     loss_f0 = f0_loss(F0_fake, F0_real)
107 |     
108 |     # style F0 loss (style initialization)
109 |     if x_refs is not None and args.lambda_f0_sty > 0 and not use_adv_cls:
110 |         F0_sty, _, _ = nets.f0_model(x_ref)
111 |         loss_f0_sty = F.l1_loss(compute_mean_f0(F0_fake), compute_mean_f0(F0_sty))
112 |     else:
113 |         loss_f0_sty = torch.zeros(1).mean()
114 |     
115 |     # ASR loss
116 |     loss_asr = F.smooth_l1_loss(ASR_fake, ASR_real)
117 |     
118 |     # style reconstruction loss
119 |     s_pred = nets.style_encoder(x_fake, y_trg)
120 |     loss_sty = torch.mean(torch.abs(s_pred - s_trg))
121 |     
122 |     # diversity sensitive loss
123 |     if z_trgs is not None:
124 |         s_trg2 = nets.mapping_network(z_trg2, y_trg)
125 |     else:
126 |         s_trg2 = nets.style_encoder(x_ref2, y_trg)
127 |     x_fake2 = nets.generator(x_real, s_trg2, masks=None, F0=GAN_F0_real)
128 |     x_fake2 = x_fake2.detach()
129 |     _, GAN_F0_fake2, _ = nets.f0_model(x_fake2)
130 |     loss_ds = torch.mean(torch.abs(x_fake - x_fake2))
131 |     loss_ds += F.smooth_l1_loss(GAN_F0_fake, GAN_F0_fake2.detach())
132 |     
133 |     # cycle-consistency loss
134 |     s_org = nets.style_encoder(x_real, y_org)
135 |     x_rec = nets.generator(x_fake, s_org, masks=None, F0=GAN_F0_fake)
136 |     loss_cyc = torch.mean(torch.abs(x_rec - x_real))
137 |     # F0 loss in cycle-consistency loss
138 |     if args.lambda_f0 > 0:
139 |         _, _, cyc_F0_rec = nets.f0_model(x_rec)
140 |         loss_cyc += F.smooth_l1_loss(cyc_F0_rec, cyc_F0_real)
141 |     if args.lambda_asr > 0:
142 |         ASR_recon = nets.asr_model.get_feature(x_rec)
143 |         loss_cyc += F.smooth_l1_loss(ASR_recon, ASR_real)
144 |     
145 |     # adversarial classifier loss
146 |     if use_adv_cls:
147 |         out_de = nets.discriminator.classifier(x_fake)
148 |         loss_adv_cls = F.cross_entropy(out_de[y_org != y_trg], y_trg[y_org != y_trg])
149 |     else:
150 |         loss_adv_cls = torch.zeros(1).mean()
151 |     
152 |     loss = args.lambda_adv * loss_adv + args.lambda_sty * loss_sty \
153 |            - args.lambda_ds * loss_ds + args.lambda_cyc * loss_cyc\
154 |            + args.lambda_norm * loss_norm \
155 |            + args.lambda_asr * loss_asr \
156 |            + args.lambda_f0 * loss_f0 \
157 |            + args.lambda_f0_sty * loss_f0_sty \
158 |            + args.lambda_adv_cls * loss_adv_cls
159 | 
160 |     return loss, Munch(adv=loss_adv.item(),
161 |                        sty=loss_sty.item(),
162 |                        ds=loss_ds.item(),
163 |                        cyc=loss_cyc.item(),
164 |                        norm=loss_norm.item(),
165 |                        asr=loss_asr.item(),
166 |                        f0=loss_f0.item(),
167 |                        adv_cls=loss_adv_cls.item())
168 |     
169 | # for norm consistency loss
170 | def log_norm(x, mean=-4, std=4, dim=2):
171 |     """
172 |     normalized log mel -> mel -> norm -> log(norm)
173 |     """
174 |     x = torch.log(torch.exp(x * std + mean).norm(dim=dim))
175 |     return x
176 | 
177 | # for adversarial loss
178 | def adv_loss(logits, target):
179 |     assert target in [1, 0]
180 |     if len(logits.shape) > 1:
181 |         logits = logits.reshape(-1)
182 |     targets = torch.full_like(logits, fill_value=target)
183 |     logits = logits.clamp(min=-10, max=10) # prevent nan
184 |     loss = F.binary_cross_entropy_with_logits(logits, targets)
185 |     return loss
186 | 
187 | # for R1 regularization loss
188 | def r1_reg(d_out, x_in):
189 |     # zero-centered gradient penalty for real images
190 |     batch_size = x_in.size(0)
191 |     grad_dout = torch.autograd.grad(
192 |         outputs=d_out.sum(), inputs=x_in,
193 |         create_graph=True, retain_graph=True, only_inputs=True
194 |     )[0]
195 |     grad_dout2 = grad_dout.pow(2)
196 |     assert(grad_dout2.size() == x_in.size())
197 |     reg = 0.5 * grad_dout2.view(batch_size, -1).sum(1).mean(0)
198 |     return reg
199 | 
200 | # for F0 consistency loss
201 | def compute_mean_f0(f0):
202 |     f0_mean = f0.mean(-1)
203 |     f0_mean = f0_mean.expand(f0.shape[-1], f0_mean.shape[0]).transpose(0, 1) # (B, M)
204 |     return f0_mean
205 | 
206 | def f0_loss(x_f0, y_f0):
207 |     """
208 |     x.shape = (B, 1, M, L): predict
209 |     y.shape = (B, 1, M, L): target
210 |     """
211 |     # compute the mean
212 |     x_mean = compute_mean_f0(x_f0)
213 |     y_mean = compute_mean_f0(y_f0)
214 |     loss = F.l1_loss(x_f0 / x_mean, y_f0 / y_mean)
215 |     return loss


--------------------------------------------------------------------------------
/meldataset.py:
--------------------------------------------------------------------------------
  1 | #coding: utf-8
  2 | 
  3 | import os
  4 | import time
  5 | import random
  6 | import random
  7 | import torch
  8 | import torchaudio
  9 | 
 10 | import numpy as np
 11 | import soundfile as sf
 12 | import torch.nn.functional as F
 13 | 
 14 | from torch import nn
 15 | from torch.utils.data import DataLoader
 16 | 
 17 | import logging
 18 | logger = logging.getLogger(__name__)
 19 | logger.setLevel(logging.DEBUG)
 20 | 
 21 | np.random.seed(1)
 22 | random.seed(1)
 23 | 
 24 | SPECT_PARAMS = {
 25 |     "n_fft": 2048,
 26 |     "win_length": 1200,
 27 |     "hop_length": 300
 28 | }
 29 | MEL_PARAMS = {
 30 |     "n_mels": 80,
 31 |     "n_fft": 2048,
 32 |     "win_length": 1200,
 33 |     "hop_length": 300
 34 | }
 35 | 
 36 | class MelDataset(torch.utils.data.Dataset):
 37 |     def __init__(self,
 38 |                  data_list,
 39 |                  sr=24000,
 40 |                  validation=False,
 41 |                  ):
 42 | 
 43 |         _data_list = [l[:-1].split('|') for l in data_list]
 44 |         self.data_list = [(path, int(label)) for path, label in _data_list]
 45 |         self.data_list_per_class = {
 46 |             target: [(path, label) for path, label in self.data_list if label == target] \
 47 |             for target in list(set([label for _, label in self.data_list]))}
 48 | 
 49 |         self.sr = sr
 50 |         self.to_melspec = torchaudio.transforms.MelSpectrogram(**MEL_PARAMS)
 51 | 
 52 |         self.mean, self.std = -4, 4
 53 |         self.validation = validation
 54 |         self.max_mel_length = 192
 55 | 
 56 |     def __len__(self):
 57 |         return len(self.data_list)
 58 | 
 59 |     def __getitem__(self, idx):
 60 |         data = self.data_list[idx]
 61 |         mel_tensor, label = self._load_data(data)
 62 |         ref_data = random.choice(self.data_list)
 63 |         ref_mel_tensor, ref_label = self._load_data(ref_data)
 64 |         ref2_data = random.choice(self.data_list_per_class[ref_label])
 65 |         ref2_mel_tensor, _ = self._load_data(ref2_data)
 66 |         return mel_tensor, label, ref_mel_tensor, ref2_mel_tensor, ref_label
 67 |     
 68 |     def _load_data(self, path):
 69 |         wave_tensor, label = self._load_tensor(path)
 70 |         
 71 |         if not self.validation: # random scale for robustness
 72 |             random_scale = 0.5 + 0.5 * np.random.random()
 73 |             wave_tensor = random_scale * wave_tensor
 74 | 
 75 |         mel_tensor = self.to_melspec(wave_tensor)
 76 |         mel_tensor = (torch.log(1e-5 + mel_tensor) - self.mean) / self.std
 77 |         mel_length = mel_tensor.size(1)
 78 |         if mel_length > self.max_mel_length:
 79 |             random_start = np.random.randint(0, mel_length - self.max_mel_length)
 80 |             mel_tensor = mel_tensor[:, random_start:random_start + self.max_mel_length]
 81 | 
 82 |         return mel_tensor, label
 83 | 
 84 |     def _preprocess(self, wave_tensor, ):
 85 |         mel_tensor = self.to_melspec(wave_tensor)
 86 |         mel_tensor = (torch.log(1e-5 + mel_tensor) - self.mean) / self.std
 87 |         return mel_tensor
 88 | 
 89 |     def _load_tensor(self, data):
 90 |         wave_path, label = data
 91 |         label = int(label)
 92 |         wave, sr = sf.read(wave_path)
 93 |         wave_tensor = torch.from_numpy(wave).float()
 94 |         return wave_tensor, label
 95 | 
 96 | class Collater(object):
 97 |     """
 98 |     Args:
 99 |       adaptive_batch_size (bool): if true, decrease batch size when long data comes.
100 |     """
101 | 
102 |     def __init__(self, return_wave=False):
103 |         self.text_pad_index = 0
104 |         self.return_wave = return_wave
105 |         self.max_mel_length = 192
106 |         self.mel_length_step = 16
107 |         self.latent_dim = 16
108 | 
109 |     def __call__(self, batch):
110 |         batch_size = len(batch)
111 |         nmels = batch[0][0].size(0)
112 |         mels = torch.zeros((batch_size, nmels, self.max_mel_length)).float()
113 |         labels = torch.zeros((batch_size)).long()
114 |         ref_mels = torch.zeros((batch_size, nmels, self.max_mel_length)).float()
115 |         ref2_mels = torch.zeros((batch_size, nmels, self.max_mel_length)).float()
116 |         ref_labels = torch.zeros((batch_size)).long()
117 | 
118 |         for bid, (mel, label, ref_mel, ref2_mel, ref_label) in enumerate(batch):
119 |             mel_size = mel.size(1)
120 |             mels[bid, :, :mel_size] = mel
121 |             
122 |             ref_mel_size = ref_mel.size(1)
123 |             ref_mels[bid, :, :ref_mel_size] = ref_mel
124 |             
125 |             ref2_mel_size = ref2_mel.size(1)
126 |             ref2_mels[bid, :, :ref2_mel_size] = ref2_mel
127 |             
128 |             labels[bid] = label
129 |             ref_labels[bid] = ref_label
130 | 
131 |         z_trg = torch.randn(batch_size, self.latent_dim)
132 |         z_trg2 = torch.randn(batch_size, self.latent_dim)
133 |         
134 |         mels, ref_mels, ref2_mels = mels.unsqueeze(1), ref_mels.unsqueeze(1), ref2_mels.unsqueeze(1)
135 |         return mels, labels, ref_mels, ref2_mels, ref_labels, z_trg, z_trg2
136 | 
137 | def build_dataloader(path_list,
138 |                      validation=False,
139 |                      batch_size=4,
140 |                      num_workers=1,
141 |                      device='cpu',
142 |                      collate_config={},
143 |                      dataset_config={}):
144 | 
145 |     dataset = MelDataset(path_list, validation=validation)
146 |     collate_fn = Collater(**collate_config)
147 |     data_loader = DataLoader(dataset,
148 |                              batch_size=batch_size,
149 |                              shuffle=(not validation),
150 |                              num_workers=num_workers,
151 |                              drop_last=True,
152 |                              collate_fn=collate_fn,
153 |                              pin_memory=(device != 'cpu'))
154 | 
155 |     return data_loader
156 | 


--------------------------------------------------------------------------------
/models.py:
--------------------------------------------------------------------------------
  1 | """
  2 | StarGAN v2
  3 | Copyright (c) 2020-present NAVER Corp.
  4 | This work is licensed under the Creative Commons Attribution-NonCommercial
  5 | 4.0 International License. To view a copy of this license, visit
  6 | http://creativecommons.org/licenses/by-nc/4.0/ or send a letter to
  7 | Creative Commons, PO Box 1866, Mountain View, CA 94042, USA.
  8 | """
  9 | import os
 10 | import os.path as osp
 11 | 
 12 | import copy
 13 | import math
 14 | 
 15 | from munch import Munch
 16 | import numpy as np
 17 | import torch
 18 | import torch.nn as nn
 19 | import torch.nn.functional as F
 20 | 
 21 | class DownSample(nn.Module):
 22 |     def __init__(self, layer_type):
 23 |         super().__init__()
 24 |         self.layer_type = layer_type
 25 | 
 26 |     def forward(self, x):
 27 |         if self.layer_type == 'none':
 28 |             return x
 29 |         elif self.layer_type == 'timepreserve':
 30 |             return F.avg_pool2d(x, (2, 1))
 31 |         elif self.layer_type == 'half':
 32 |             return F.avg_pool2d(x, 2)
 33 |         else:
 34 |             raise RuntimeError('Got unexpected donwsampletype %s, expected is [none, timepreserve, half]' % self.layer_type)
 35 | 
 36 | 
 37 | class UpSample(nn.Module):
 38 |     def __init__(self, layer_type):
 39 |         super().__init__()
 40 |         self.layer_type = layer_type
 41 | 
 42 |     def forward(self, x):
 43 |         if self.layer_type == 'none':
 44 |             return x
 45 |         elif self.layer_type == 'timepreserve':
 46 |             return F.interpolate(x, scale_factor=(2, 1), mode='nearest')
 47 |         elif self.layer_type == 'half':
 48 |             return F.interpolate(x, scale_factor=2, mode='nearest')
 49 |         else:
 50 |             raise RuntimeError('Got unexpected upsampletype %s, expected is [none, timepreserve, half]' % self.layer_type)
 51 | 
 52 | 
 53 | class ResBlk(nn.Module):
 54 |     def __init__(self, dim_in, dim_out, actv=nn.LeakyReLU(0.2),
 55 |                  normalize=False, downsample='none'):
 56 |         super().__init__()
 57 |         self.actv = actv
 58 |         self.normalize = normalize
 59 |         self.downsample = DownSample(downsample)
 60 |         self.learned_sc = dim_in != dim_out
 61 |         self._build_weights(dim_in, dim_out)
 62 | 
 63 |     def _build_weights(self, dim_in, dim_out):
 64 |         self.conv1 = nn.Conv2d(dim_in, dim_in, 3, 1, 1)
 65 |         self.conv2 = nn.Conv2d(dim_in, dim_out, 3, 1, 1)
 66 |         if self.normalize:
 67 |             self.norm1 = nn.InstanceNorm2d(dim_in, affine=True)
 68 |             self.norm2 = nn.InstanceNorm2d(dim_in, affine=True)
 69 |         if self.learned_sc:
 70 |             self.conv1x1 = nn.Conv2d(dim_in, dim_out, 1, 1, 0, bias=False)
 71 | 
 72 |     def _shortcut(self, x):
 73 |         if self.learned_sc:
 74 |             x = self.conv1x1(x)
 75 |         if self.downsample:
 76 |             x = self.downsample(x)
 77 |         return x
 78 | 
 79 |     def _residual(self, x):
 80 |         if self.normalize:
 81 |             x = self.norm1(x)
 82 |         x = self.actv(x)
 83 |         x = self.conv1(x)
 84 |         x = self.downsample(x)
 85 |         if self.normalize:
 86 |             x = self.norm2(x)
 87 |         x = self.actv(x)
 88 |         x = self.conv2(x)
 89 |         return x
 90 | 
 91 |     def forward(self, x):
 92 |         x = self._shortcut(x) + self._residual(x)
 93 |         return x / math.sqrt(2)  # unit variance
 94 | 
 95 | class AdaIN(nn.Module):
 96 |     def __init__(self, style_dim, num_features):
 97 |         super().__init__()
 98 |         self.norm = nn.InstanceNorm2d(num_features, affine=False)
 99 |         self.fc = nn.Linear(style_dim, num_features*2)
100 | 
101 |     def forward(self, x, s):
102 |         h = self.fc(s)
103 |         h = h.view(h.size(0), h.size(1), 1, 1)
104 |         gamma, beta = torch.chunk(h, chunks=2, dim=1)
105 |         return (1 + gamma) * self.norm(x) + beta
106 | 
107 | 
108 | class AdainResBlk(nn.Module):
109 |     def __init__(self, dim_in, dim_out, style_dim=64, w_hpf=0,
110 |                  actv=nn.LeakyReLU(0.2), upsample='none'):
111 |         super().__init__()
112 |         self.w_hpf = w_hpf
113 |         self.actv = actv
114 |         self.upsample = UpSample(upsample)
115 |         self.learned_sc = dim_in != dim_out
116 |         self._build_weights(dim_in, dim_out, style_dim)
117 | 
118 |     def _build_weights(self, dim_in, dim_out, style_dim=64):
119 |         self.conv1 = nn.Conv2d(dim_in, dim_out, 3, 1, 1)
120 |         self.conv2 = nn.Conv2d(dim_out, dim_out, 3, 1, 1)
121 |         self.norm1 = AdaIN(style_dim, dim_in)
122 |         self.norm2 = AdaIN(style_dim, dim_out)
123 |         if self.learned_sc:
124 |             self.conv1x1 = nn.Conv2d(dim_in, dim_out, 1, 1, 0, bias=False)
125 | 
126 |     def _shortcut(self, x):
127 |         x = self.upsample(x)
128 |         if self.learned_sc:
129 |             x = self.conv1x1(x)
130 |         return x
131 | 
132 |     def _residual(self, x, s):
133 |         x = self.norm1(x, s)
134 |         x = self.actv(x)
135 |         x = self.upsample(x)
136 |         x = self.conv1(x)
137 |         x = self.norm2(x, s)
138 |         x = self.actv(x)
139 |         x = self.conv2(x)
140 |         return x
141 | 
142 |     def forward(self, x, s):
143 |         out = self._residual(x, s)
144 |         if self.w_hpf == 0:
145 |             out = (out + self._shortcut(x)) / math.sqrt(2)
146 |         return out
147 | 
148 | 
149 | class HighPass(nn.Module):
150 |     def __init__(self, w_hpf, device):
151 |         super(HighPass, self).__init__()
152 |         self.filter = torch.tensor([[-1, -1, -1],
153 |                                     [-1, 8., -1],
154 |                                     [-1, -1, -1]]).to(device) / w_hpf
155 | 
156 |     def forward(self, x):
157 |         filter = self.filter.unsqueeze(0).unsqueeze(1).repeat(x.size(1), 1, 1, 1)
158 |         return F.conv2d(x, filter, padding=1, groups=x.size(1))
159 | 
160 | 
161 | class Generator(nn.Module):
162 |     def __init__(self, dim_in=48, style_dim=48, max_conv_dim=48*8, w_hpf=1, F0_channel=0):
163 |         super().__init__()
164 | 
165 |         self.stem = nn.Conv2d(1, dim_in, 3, 1, 1)
166 |         self.encode = nn.ModuleList()
167 |         self.decode = nn.ModuleList()
168 |         self.to_out = nn.Sequential(
169 |             nn.InstanceNorm2d(dim_in, affine=True),
170 |             nn.LeakyReLU(0.2),
171 |             nn.Conv2d(dim_in, 1, 1, 1, 0))
172 |         self.F0_channel = F0_channel
173 |         # down/up-sampling blocks
174 |         repeat_num = 4 #int(np.log2(img_size)) - 4
175 |         if w_hpf > 0:
176 |             repeat_num += 1
177 | 
178 |         for lid in range(repeat_num):
179 |             if lid in [1, 3]:
180 |                 _downtype = 'timepreserve'
181 |             else:
182 |                 _downtype = 'half'
183 | 
184 |             dim_out = min(dim_in*2, max_conv_dim)
185 |             self.encode.append(
186 |                 ResBlk(dim_in, dim_out, normalize=True, downsample=_downtype))
187 |             self.decode.insert(
188 |                 0, AdainResBlk(dim_out, dim_in, style_dim,
189 |                                w_hpf=w_hpf, upsample=_downtype))  # stack-like
190 |             dim_in = dim_out
191 | 
192 |         # bottleneck blocks (encoder)
193 |         for _ in range(2):
194 |             self.encode.append(
195 |                 ResBlk(dim_out, dim_out, normalize=True))
196 |         
197 |         # F0 blocks 
198 |         if F0_channel != 0:
199 |             self.decode.insert(
200 |                 0, AdainResBlk(dim_out + int(F0_channel / 2), dim_out, style_dim, w_hpf=w_hpf))
201 |         
202 |         # bottleneck blocks (decoder)
203 |         for _ in range(2):
204 |             self.decode.insert(
205 |                     0, AdainResBlk(dim_out + int(F0_channel / 2), dim_out + int(F0_channel / 2), style_dim, w_hpf=w_hpf))
206 |         
207 |         if F0_channel != 0:
208 |             self.F0_conv = nn.Sequential(
209 |                 ResBlk(F0_channel, int(F0_channel / 2), normalize=True, downsample="half"),
210 |             )
211 |         
212 | 
213 |         if w_hpf > 0:
214 |             device = torch.device(
215 |                 'cuda' if torch.cuda.is_available() else 'cpu')
216 |             self.hpf = HighPass(w_hpf, device)
217 | 
218 |     def forward(self, x, s, masks=None, F0=None):            
219 |         x = self.stem(x)
220 |         cache = {}
221 |         for block in self.encode:
222 |             if (masks is not None) and (x.size(2) in [32, 64, 128]):
223 |                 cache[x.size(2)] = x
224 |             x = block(x)
225 |             
226 |         if F0 is not None:
227 |             F0 = self.F0_conv(F0)
228 |             F0 = F.adaptive_avg_pool2d(F0, [x.shape[-2], x.shape[-1]])
229 |             x = torch.cat([x, F0], axis=1)
230 | 
231 |         for block in self.decode:
232 |             x = block(x, s)
233 |             if (masks is not None) and (x.size(2) in [32, 64, 128]):
234 |                 mask = masks[0] if x.size(2) in [32] else masks[1]
235 |                 mask = F.interpolate(mask, size=x.size(2), mode='bilinear')
236 |                 x = x + self.hpf(mask * cache[x.size(2)])
237 | 
238 |         return self.to_out(x)
239 | 
240 | 
241 | class MappingNetwork(nn.Module):
242 |     def __init__(self, latent_dim=16, style_dim=48, num_domains=2, hidden_dim=384):
243 |         super().__init__()
244 |         layers = []
245 |         layers += [nn.Linear(latent_dim, hidden_dim)]
246 |         layers += [nn.ReLU()]
247 |         for _ in range(3):
248 |             layers += [nn.Linear(hidden_dim, hidden_dim)]
249 |             layers += [nn.ReLU()]
250 |         self.shared = nn.Sequential(*layers)
251 | 
252 |         self.unshared = nn.ModuleList()
253 |         for _ in range(num_domains):
254 |             self.unshared += [nn.Sequential(nn.Linear(hidden_dim, hidden_dim),
255 |                                             nn.ReLU(),
256 |                                             nn.Linear(hidden_dim, hidden_dim),
257 |                                             nn.ReLU(),
258 |                                             nn.Linear(hidden_dim, hidden_dim),
259 |                                             nn.ReLU(),
260 |                                             nn.Linear(hidden_dim, style_dim))]
261 | 
262 |     def forward(self, z, y):
263 |         h = self.shared(z)
264 |         out = []
265 |         for layer in self.unshared:
266 |             out += [layer(h)]
267 |         out = torch.stack(out, dim=1)  # (batch, num_domains, style_dim)
268 |         idx = torch.LongTensor(range(y.size(0))).to(y.device)
269 |         s = out[idx, y]  # (batch, style_dim)
270 |         return s
271 | 
272 | 
273 | class StyleEncoder(nn.Module):
274 |     def __init__(self, dim_in=48, style_dim=48, num_domains=2, max_conv_dim=384):
275 |         super().__init__()
276 |         blocks = []
277 |         blocks += [nn.Conv2d(1, dim_in, 3, 1, 1)]
278 | 
279 |         repeat_num = 4
280 |         for _ in range(repeat_num):
281 |             dim_out = min(dim_in*2, max_conv_dim)
282 |             blocks += [ResBlk(dim_in, dim_out, downsample='half')]
283 |             dim_in = dim_out
284 | 
285 |         blocks += [nn.LeakyReLU(0.2)]
286 |         blocks += [nn.Conv2d(dim_out, dim_out, 5, 1, 0)]
287 |         blocks += [nn.AdaptiveAvgPool2d(1)]
288 |         blocks += [nn.LeakyReLU(0.2)]
289 |         self.shared = nn.Sequential(*blocks)
290 | 
291 |         self.unshared = nn.ModuleList()
292 |         for _ in range(num_domains):
293 |             self.unshared += [nn.Linear(dim_out, style_dim)]
294 | 
295 |     def forward(self, x, y):
296 |         h = self.shared(x)
297 | 
298 |         h = h.view(h.size(0), -1)
299 |         out = []
300 | 
301 |         for layer in self.unshared:
302 |             out += [layer(h)]
303 | 
304 |         out = torch.stack(out, dim=1)  # (batch, num_domains, style_dim)
305 |         idx = torch.LongTensor(range(y.size(0))).to(y.device)
306 |         s = out[idx, y]  # (batch, style_dim)
307 |         return s
308 | 
309 | class Discriminator(nn.Module):
310 |     def __init__(self, dim_in=48, num_domains=2, max_conv_dim=384, repeat_num=4):
311 |         super().__init__()
312 |         
313 |         # real/fake discriminator
314 |         self.dis = Discriminator2d(dim_in=dim_in, num_domains=num_domains,
315 |                                   max_conv_dim=max_conv_dim, repeat_num=repeat_num)
316 |         # adversarial classifier
317 |         self.cls = Discriminator2d(dim_in=dim_in, num_domains=num_domains,
318 |                                   max_conv_dim=max_conv_dim, repeat_num=repeat_num)                             
319 |         self.num_domains = num_domains
320 |         
321 |     def forward(self, x, y):
322 |         return self.dis(x, y)
323 | 
324 |     def classifier(self, x):
325 |         return self.cls.get_feature(x)
326 | 
327 | 
328 | class LinearNorm(torch.nn.Module):
329 |     def __init__(self, in_dim, out_dim, bias=True, w_init_gain='linear'):
330 |         super(LinearNorm, self).__init__()
331 |         self.linear_layer = torch.nn.Linear(in_dim, out_dim, bias=bias)
332 | 
333 |         torch.nn.init.xavier_uniform_(
334 |             self.linear_layer.weight,
335 |             gain=torch.nn.init.calculate_gain(w_init_gain))
336 | 
337 |     def forward(self, x):
338 |         return self.linear_layer(x)
339 | 
340 | class Discriminator2d(nn.Module):
341 |     def __init__(self, dim_in=48, num_domains=2, max_conv_dim=384, repeat_num=4):
342 |         super().__init__()
343 |         blocks = []
344 |         blocks += [nn.Conv2d(1, dim_in, 3, 1, 1)]
345 | 
346 |         for lid in range(repeat_num):
347 |             dim_out = min(dim_in*2, max_conv_dim)
348 |             blocks += [ResBlk(dim_in, dim_out, downsample='half')]
349 |             dim_in = dim_out
350 | 
351 |         blocks += [nn.LeakyReLU(0.2)]
352 |         blocks += [nn.Conv2d(dim_out, dim_out, 5, 1, 0)]
353 |         blocks += [nn.LeakyReLU(0.2)]
354 |         blocks += [nn.AdaptiveAvgPool2d(1)]
355 |         blocks += [nn.Conv2d(dim_out, num_domains, 1, 1, 0)]
356 |         self.main = nn.Sequential(*blocks)
357 | 
358 |     def get_feature(self, x):
359 |         out = self.main(x)
360 |         out = out.view(out.size(0), -1)  # (batch, num_domains)
361 |         return out
362 | 
363 |     def forward(self, x, y):
364 |         out = self.get_feature(x)
365 |         idx = torch.LongTensor(range(y.size(0))).to(y.device)
366 |         out = out[idx, y]  # (batch)
367 |         return out
368 | 
369 | 
370 | def build_model(args, F0_model, ASR_model):
371 |     generator = Generator(args.dim_in, args.style_dim, args.max_conv_dim, w_hpf=args.w_hpf, F0_channel=args.F0_channel)
372 |     mapping_network = MappingNetwork(args.latent_dim, args.style_dim, args.num_domains, hidden_dim=args.max_conv_dim)
373 |     style_encoder = StyleEncoder(args.dim_in, args.style_dim, args.num_domains, args.max_conv_dim)
374 |     discriminator = Discriminator(args.dim_in, args.num_domains, args.max_conv_dim, args.n_repeat)
375 |     generator_ema = copy.deepcopy(generator)
376 |     mapping_network_ema = copy.deepcopy(mapping_network)
377 |     style_encoder_ema = copy.deepcopy(style_encoder)
378 |         
379 |     nets = Munch(generator=generator,
380 |                  mapping_network=mapping_network,
381 |                  style_encoder=style_encoder,
382 |                  discriminator=discriminator,
383 |                  f0_model=F0_model,
384 |                  asr_model=ASR_model)
385 |     
386 |     nets_ema = Munch(generator=generator_ema,
387 |                      mapping_network=mapping_network_ema,
388 |                      style_encoder=style_encoder_ema)
389 | 
390 |     return nets, nets_ema


--------------------------------------------------------------------------------
/optimizers.py:
--------------------------------------------------------------------------------
 1 | #coding:utf-8
 2 | import os, sys
 3 | import os.path as osp
 4 | import numpy as np
 5 | import torch
 6 | from torch import nn
 7 | from torch.optim import Optimizer
 8 | from functools import reduce
 9 | from torch.optim import AdamW
10 | 
11 | class MultiOptimizer:
12 |     def __init__(self, optimizers={}, schedulers={}):
13 |         self.optimizers = optimizers
14 |         self.schedulers = schedulers
15 |         self.keys = list(optimizers.keys())
16 |         self.param_groups = reduce(lambda x,y: x+y, [v.param_groups for v in self.optimizers.values()])
17 | 
18 |     def state_dict(self):
19 |         state_dicts = [(key, self.optimizers[key].state_dict())\
20 |                        for key in self.keys]
21 |         return state_dicts
22 | 
23 |     def load_state_dict(self, state_dict):
24 |         for key, val in state_dict:
25 |             try:
26 |                 self.optimizers[key].load_state_dict(val)
27 |             except:
28 |                 print("Unloaded %s" % key)
29 | 
30 |     def step(self, key=None, scaler=None):
31 |         keys = [key] if key is not None else self.keys
32 |         _ = [self._step(key, scaler) for key in keys]
33 | 
34 |     def _step(self, key, scaler=None):
35 |         if scaler is not None:
36 |             scaler.step(self.optimizers[key])
37 |             scaler.update()
38 |         else:
39 |             self.optimizers[key].step()
40 | 
41 |     def zero_grad(self, key=None):
42 |         if key is not None:
43 |             self.optimizers[key].zero_grad()
44 |         else:
45 |             _ = [self.optimizers[key].zero_grad() for key in self.keys]
46 | 
47 |     def scheduler(self, *args, key=None):
48 |         if key is not None:
49 |             self.schedulers[key].step(*args)
50 |         else:
51 |             _ = [self.schedulers[key].step(*args) for key in self.keys]
52 | 
53 | def define_scheduler(optimizer, params):
54 |     print(params)
55 |     scheduler = torch.optim.lr_scheduler.OneCycleLR(
56 |         optimizer,
57 |         max_lr=params.get('max_lr', 2e-4),
58 |         epochs=params.get('epochs', 200),
59 |         steps_per_epoch=params.get('steps_per_epoch', 1000),
60 |         pct_start=params.get('pct_start', 0.0),
61 |         div_factor=1,
62 |         final_div_factor=1)
63 | 
64 |     return scheduler
65 | 
66 | def build_optimizer(parameters_dict, scheduler_params_dict):
67 |     optim = dict([(key, AdamW(params, lr=1e-4, weight_decay=1e-4, betas=(0.0, 0.99), eps=1e-9))
68 |                    for key, params in parameters_dict.items()])
69 | 
70 |     schedulers = dict([(key, define_scheduler(opt, scheduler_params_dict[key])) \
71 |                        for key, opt in optim.items()])
72 | 
73 |     multi_optim = MultiOptimizer(optim, schedulers)
74 |     return multi_optim


--------------------------------------------------------------------------------
/train.py:
--------------------------------------------------------------------------------
  1 | #!/usr/bin/env python3
  2 | #coding:utf-8
  3 | 
  4 | import os
  5 | import os.path as osp
  6 | import re
  7 | import sys
  8 | import yaml
  9 | import shutil
 10 | import numpy as np
 11 | import torch
 12 | import click
 13 | import warnings
 14 | warnings.simplefilter('ignore')
 15 | 
 16 | from functools import reduce
 17 | from munch import Munch
 18 | 
 19 | from meldataset import build_dataloader
 20 | from optimizers import build_optimizer
 21 | from models import build_model
 22 | from trainer import Trainer
 23 | from torch.utils.tensorboard import SummaryWriter
 24 | 
 25 | from Utils.ASR.models import ASRCNN
 26 | from Utils.JDC.model import JDCNet
 27 | 
 28 | import logging
 29 | from logging import StreamHandler
 30 | logger = logging.getLogger(__name__)
 31 | logger.setLevel(logging.DEBUG)
 32 | handler = StreamHandler()
 33 | handler.setLevel(logging.DEBUG)
 34 | logger.addHandler(handler)
 35 | 
 36 | torch.backends.cudnn.benchmark = True #
 37 | 
 38 | @click.command()
 39 | @click.option('-p', '--config_path', default='Configs/config.yml', type=str)
 40 | 
 41 | def main(config_path):
 42 |     config = yaml.safe_load(open(config_path))
 43 | 
 44 |     log_dir = config['log_dir']
 45 |     if not osp.exists(log_dir): os.makedirs(log_dir, exist_ok=True)
 46 |     shutil.copy(config_path, osp.join(log_dir, osp.basename(config_path)))
 47 |     writer = SummaryWriter(log_dir + "/tensorboard")
 48 | 
 49 |     # write logs
 50 |     file_handler = logging.FileHandler(osp.join(log_dir, 'train.log'))
 51 |     file_handler.setLevel(logging.DEBUG)
 52 |     file_handler.setFormatter(logging.Formatter('%(levelname)s:%(asctime)s: %(message)s'))
 53 |     logger.addHandler(file_handler)
 54 |     
 55 |     batch_size = config.get('batch_size', 10)
 56 |     device = config.get('device', 'cpu')
 57 |     epochs = config.get('epochs', 1000)
 58 |     save_freq = config.get('save_freq', 20)
 59 |     train_path = config.get('train_data', None)
 60 |     val_path = config.get('val_data', None)
 61 |     stage = config.get('stage', 'star')
 62 |     fp16_run = config.get('fp16_run', False)
 63 |     
 64 |     # load data
 65 |     train_list, val_list = get_data_path_list(train_path, val_path)
 66 |     train_dataloader = build_dataloader(train_list,
 67 |                                         batch_size=batch_size,
 68 |                                         num_workers=4,
 69 |                                         device=device)
 70 |     val_dataloader = build_dataloader(val_list,
 71 |                                       batch_size=batch_size,
 72 |                                       validation=True,
 73 |                                       num_workers=2,
 74 |                                       device=device)
 75 | 
 76 |     # load pretrained ASR model
 77 |     ASR_config = config.get('ASR_config', False)
 78 |     ASR_path = config.get('ASR_path', False)
 79 |     with open(ASR_config) as f:
 80 |             ASR_config = yaml.safe_load(f)
 81 |     ASR_model_config = ASR_config['model_params']
 82 |     ASR_model = ASRCNN(**ASR_model_config)
 83 |     params = torch.load(ASR_path, map_location='cpu')['model']
 84 |     ASR_model.load_state_dict(params)
 85 |     _ = ASR_model.eval()
 86 |     
 87 |     # load pretrained F0 model
 88 |     F0_path = config.get('F0_path', False)
 89 |     F0_model = JDCNet(num_class=1, seq_len=192)
 90 |     params = torch.load(F0_path, map_location='cpu')['net']
 91 |     F0_model.load_state_dict(params)
 92 |     
 93 |     # build model
 94 |     model, model_ema = build_model(Munch(config['model_params']), F0_model, ASR_model)
 95 | 
 96 |     scheduler_params = {
 97 |         "max_lr": float(config['optimizer_params'].get('lr', 2e-4)),
 98 |         "pct_start": float(config['optimizer_params'].get('pct_start', 0.0)),
 99 |         "epochs": epochs,
100 |         "steps_per_epoch": len(train_dataloader),
101 |     }
102 |     
103 |     _ = [model[key].to(device) for key in model]
104 |     _ = [model_ema[key].to(device) for key in model_ema]
105 |     scheduler_params_dict = {key: scheduler_params.copy() for key in model}
106 |     scheduler_params_dict['mapping_network']['max_lr'] = 2e-6
107 |     optimizer = build_optimizer({key: model[key].parameters() for key in model},
108 |                                       scheduler_params_dict=scheduler_params_dict)
109 | 
110 |     trainer = Trainer(args=Munch(config['loss_params']), model=model,
111 |                             model_ema=model_ema,
112 |                             optimizer=optimizer,
113 |                             device=device,
114 |                             train_dataloader=train_dataloader,
115 |                             val_dataloader=val_dataloader,
116 |                             logger=logger,
117 |                             fp16_run=fp16_run)
118 | 
119 |     if config.get('pretrained_model', '') != '':
120 |         trainer.load_checkpoint(config['pretrained_model'],
121 |                                 load_only_params=config.get('load_only_params', True))
122 | 
123 |     for _ in range(1, epochs+1):
124 |         epoch = trainer.epochs
125 |         train_results = trainer._train_epoch()
126 |         eval_results = trainer._eval_epoch()
127 |         results = train_results.copy()
128 |         results.update(eval_results)
129 |         logger.info('--- epoch %d ---' % epoch)
130 |         for key, value in results.items():
131 |             if isinstance(value, float):
132 |                 logger.info('%-15s: %.4f' % (key, value))
133 |                 writer.add_scalar(key, value, epoch)
134 |             else:
135 |                 for v in value:
136 |                     writer.add_figure('eval_spec', v, epoch)
137 |         if (epoch % save_freq) == 0:
138 |             trainer.save_checkpoint(osp.join(log_dir, 'epoch_%05d.pth' % epoch))
139 | 
140 |     return 0
141 | 
142 | def get_data_path_list(train_path=None, val_path=None):
143 |     if train_path is None:
144 |         train_path = "Data/train_list.txt"
145 |     if val_path is None:
146 |         val_path = "Data/val_list.txt"
147 | 
148 |     with open(train_path, 'r') as f:
149 |         train_list = f.readlines()
150 |     with open(val_path, 'r') as f:
151 |         val_list = f.readlines()
152 | 
153 |     return train_list, val_list
154 | 
155 | if __name__=="__main__":
156 |     main()
157 | 


--------------------------------------------------------------------------------
/trainer.py:
--------------------------------------------------------------------------------
  1 | # -*- coding: utf-8 -*-
  2 | 
  3 | import os
  4 | import os.path as osp
  5 | import sys
  6 | import time
  7 | from collections import defaultdict
  8 | 
  9 | import numpy as np
 10 | import torch
 11 | from torch import nn
 12 | from PIL import Image
 13 | from tqdm import tqdm
 14 | 
 15 | from losses import compute_d_loss, compute_g_loss
 16 | 
 17 | import logging
 18 | logger = logging.getLogger(__name__)
 19 | logger.setLevel(logging.DEBUG)
 20 | 
 21 | class Trainer(object):
 22 |     def __init__(self,
 23 |                  args,
 24 |                  model=None,
 25 |                  model_ema=None,
 26 |                  optimizer=None,
 27 |                  scheduler=None,
 28 |                  config={},
 29 |                  device=torch.device("cpu"),
 30 |                  logger=logger,
 31 |                  train_dataloader=None,
 32 |                  val_dataloader=None,
 33 |                  initial_steps=0,
 34 |                  initial_epochs=0,
 35 |                  fp16_run=False
 36 |     ):
 37 |         self.args = args
 38 |         self.steps = initial_steps
 39 |         self.epochs = initial_epochs
 40 |         self.model = model
 41 |         self.model_ema = model_ema
 42 |         self.optimizer = optimizer
 43 |         self.scheduler = scheduler
 44 |         self.train_dataloader = train_dataloader
 45 |         self.val_dataloader = val_dataloader
 46 |         self.config = config
 47 |         self.device = device
 48 |         self.finish_train = False
 49 |         self.logger = logger
 50 |         self.fp16_run = fp16_run
 51 | 
 52 |     def _train_epoch(self):
 53 |         """Train model one epoch."""
 54 |         raise NotImplementedError
 55 | 
 56 |     @torch.no_grad()
 57 |     def _eval_epoch(self):
 58 |         """Evaluate model one epoch."""
 59 |         pass
 60 | 
 61 |     def save_checkpoint(self, checkpoint_path):
 62 |         """Save checkpoint.
 63 |         Args:
 64 |             checkpoint_path (str): Checkpoint path to be saved.
 65 |         """
 66 |         state_dict = {
 67 |             "optimizer": self.optimizer.state_dict(),
 68 |             "steps": self.steps,
 69 |             "epochs": self.epochs,
 70 |             "model": {key: self.model[key].state_dict() for key in self.model}
 71 |         }
 72 |         if self.model_ema is not None:
 73 |             state_dict['model_ema'] = {key: self.model_ema[key].state_dict() for key in self.model_ema}
 74 | 
 75 |         if not os.path.exists(os.path.dirname(checkpoint_path)):
 76 |             os.makedirs(os.path.dirname(checkpoint_path))
 77 |         torch.save(state_dict, checkpoint_path)
 78 | 
 79 |     def load_checkpoint(self, checkpoint_path, load_only_params=False):
 80 |         """Load checkpoint.
 81 | 
 82 |         Args:
 83 |             checkpoint_path (str): Checkpoint path to be loaded.
 84 |             load_only_params (bool): Whether to load only model parameters.
 85 | 
 86 |         """
 87 |         state_dict = torch.load(checkpoint_path, map_location="cpu")
 88 |         for key in self.model:
 89 |             self._load(state_dict["model"][key], self.model[key])
 90 | 
 91 |         if self.model_ema is not None:
 92 |             for key in self.model_ema:
 93 |                 self._load(state_dict["model_ema"][key], self.model_ema[key])
 94 |         
 95 |         if not load_only_params:
 96 |             self.steps = state_dict["steps"]
 97 |             self.epochs = state_dict["epochs"]
 98 |             self.optimizer.load_state_dict(state_dict["optimizer"])
 99 | 
100 | 
101 |     def _load(self, states, model, force_load=True):
102 |         model_states = model.state_dict()
103 |         for key, val in states.items():
104 |             try:
105 |                 if key not in model_states:
106 |                     continue
107 |                 if isinstance(val, nn.Parameter):
108 |                     val = val.data
109 | 
110 |                 if val.shape != model_states[key].shape:
111 |                     self.logger.info("%s does not have same shape" % key)
112 |                     print(val.shape, model_states[key].shape)
113 |                     if not force_load:
114 |                         continue
115 | 
116 |                     min_shape = np.minimum(np.array(val.shape), np.array(model_states[key].shape))
117 |                     slices = [slice(0, min_index) for min_index in min_shape]
118 |                     model_states[key][slices].copy_(val[slices])
119 |                 else:
120 |                     model_states[key].copy_(val)
121 |             except:
122 |                 self.logger.info("not exist :%s" % key)
123 |                 print("not exist ", key)
124 | 
125 |     @staticmethod
126 |     def get_gradient_norm(model):
127 |         total_norm = 0
128 |         for p in model.parameters():
129 |             param_norm = p.grad.data.norm(2)
130 |             total_norm += param_norm.item() ** 2
131 | 
132 |         total_norm = np.sqrt(total_norm)
133 |         return total_norm
134 | 
135 |     @staticmethod
136 |     def length_to_mask(lengths):
137 |         mask = torch.arange(lengths.max()).unsqueeze(0).expand(lengths.shape[0], -1).type_as(lengths)
138 |         mask = torch.gt(mask+1, lengths.unsqueeze(1))
139 |         return mask
140 | 
141 |     def _get_lr(self):
142 |         for param_group in self.optimizer.param_groups:
143 |             lr = param_group['lr']
144 |             break
145 |         return lr
146 | 
147 |     @staticmethod
148 |     def moving_average(model, model_test, beta=0.999):
149 |         for param, param_test in zip(model.parameters(), model_test.parameters()):
150 |             param_test.data = torch.lerp(param.data, param_test.data, beta)
151 | 
152 |     def _train_epoch(self):
153 |         self.epochs += 1
154 |         
155 |         train_losses = defaultdict(list)
156 |         _ = [self.model[k].train() for k in self.model]
157 |         scaler = torch.cuda.amp.GradScaler() if (('cuda' in str(self.device)) and self.fp16_run) else None
158 | 
159 |         use_con_reg = (self.epochs >= self.args.con_reg_epoch)
160 |         use_adv_cls = (self.epochs >= self.args.adv_cls_epoch)
161 |         
162 |         for train_steps_per_epoch, batch in enumerate(tqdm(self.train_dataloader, desc="[train]"), 1):
163 | 
164 |             ### load data
165 |             batch = [b.to(self.device) for b in batch]
166 |             x_real, y_org, x_ref, x_ref2, y_trg, z_trg, z_trg2 = batch
167 |             
168 |             # train the discriminator (by random reference)
169 |             self.optimizer.zero_grad()
170 |             if scaler is not None:
171 |                 with torch.cuda.amp.autocast():
172 |                     d_loss, d_losses_latent = compute_d_loss(self.model, self.args.d_loss, x_real, y_org, y_trg, z_trg=z_trg, use_adv_cls=use_adv_cls, use_con_reg=use_con_reg)
173 |                 scaler.scale(d_loss).backward()
174 |             else:
175 |                 d_loss, d_losses_latent = compute_d_loss(self.model, self.args.d_loss, x_real, y_org, y_trg, z_trg=z_trg, use_adv_cls=use_adv_cls, use_con_reg=use_con_reg)
176 |                 d_loss.backward()
177 |             self.optimizer.step('discriminator', scaler=scaler)
178 | 
179 |             # train the discriminator (by target reference)
180 |             self.optimizer.zero_grad()
181 |             if scaler is not None:
182 |                 with torch.cuda.amp.autocast():
183 |                     d_loss, d_losses_ref = compute_d_loss(self.model, self.args.d_loss, x_real, y_org, y_trg, x_ref=x_ref, use_adv_cls=use_adv_cls, use_con_reg=use_con_reg)
184 |                 scaler.scale(d_loss).backward()
185 |             else:
186 |                 d_loss, d_losses_ref = compute_d_loss(self.model, self.args.d_loss, x_real, y_org, y_trg, x_ref=x_ref, use_adv_cls=use_adv_cls, use_con_reg=use_con_reg)
187 |                 d_loss.backward()
188 | 
189 |             self.optimizer.step('discriminator', scaler=scaler)
190 | 
191 |             # train the generator (by random reference)
192 |             self.optimizer.zero_grad()
193 |             if scaler is not None:
194 |                 with torch.cuda.amp.autocast():
195 |                     g_loss, g_losses_latent = compute_g_loss(
196 |                         self.model, self.args.g_loss, x_real, y_org, y_trg, z_trgs=[z_trg, z_trg2], use_adv_cls=use_adv_cls)
197 |                 scaler.scale(g_loss).backward()
198 |             else:
199 |                 g_loss, g_losses_latent = compute_g_loss(
200 |                     self.model, self.args.g_loss, x_real, y_org, y_trg, z_trgs=[z_trg, z_trg2], use_adv_cls=use_adv_cls)
201 |                 g_loss.backward()
202 | 
203 |             self.optimizer.step('generator', scaler=scaler)
204 |             self.optimizer.step('mapping_network', scaler=scaler)
205 |             self.optimizer.step('style_encoder', scaler=scaler)
206 | 
207 |             # train the generator (by target reference)
208 |             self.optimizer.zero_grad()
209 |             if scaler is not None:
210 |                 with torch.cuda.amp.autocast():
211 |                     g_loss, g_losses_ref = compute_g_loss(
212 |                         self.model, self.args.g_loss, x_real, y_org, y_trg, x_refs=[x_ref, x_ref2], use_adv_cls=use_adv_cls)
213 |                 scaler.scale(g_loss).backward()
214 |             else:
215 |                 g_loss, g_losses_ref = compute_g_loss(
216 |                     self.model, self.args.g_loss, x_real, y_org, y_trg, x_refs=[x_ref, x_ref2], use_adv_cls=use_adv_cls)
217 |                 g_loss.backward()
218 |             self.optimizer.step('generator', scaler=scaler)
219 | 
220 |             # compute moving average of network parameters
221 |             self.moving_average(self.model.generator, self.model_ema.generator, beta=0.999)
222 |             self.moving_average(self.model.mapping_network, self.model_ema.mapping_network, beta=0.999)
223 |             self.moving_average(self.model.style_encoder, self.model_ema.style_encoder, beta=0.999)
224 |             self.optimizer.scheduler()
225 | 
226 |             for key in d_losses_latent:
227 |                 train_losses["train/%s" % key].append(d_losses_latent[key])
228 |             for key in g_losses_latent:
229 |                 train_losses["train/%s" % key].append(g_losses_latent[key])
230 | 
231 | 
232 |         train_losses = {key: np.mean(value) for key, value in train_losses.items()}
233 |         return train_losses
234 | 
235 |     @torch.no_grad()
236 |     def _eval_epoch(self):
237 |         use_adv_cls = (self.epochs >= self.args.adv_cls_epoch)
238 |         
239 |         eval_losses = defaultdict(list)
240 |         eval_images = defaultdict(list)
241 |         _ = [self.model[k].eval() for k in self.model]
242 |         for eval_steps_per_epoch, batch in enumerate(tqdm(self.val_dataloader, desc="[eval]"), 1):
243 | 
244 |             ### load data
245 |             batch = [b.to(self.device) for b in batch]
246 |             x_real, y_org, x_ref, x_ref2, y_trg, z_trg, z_trg2 = batch
247 | 
248 |             # train the discriminator
249 |             d_loss, d_losses_latent = compute_d_loss(
250 |                 self.model, self.args.d_loss, x_real, y_org, y_trg, z_trg=z_trg, use_r1_reg=False, use_adv_cls=use_adv_cls)
251 |             d_loss, d_losses_ref = compute_d_loss(
252 |                 self.model, self.args.d_loss, x_real, y_org, y_trg, x_ref=x_ref, use_r1_reg=False, use_adv_cls=use_adv_cls)
253 | 
254 |             # train the generator
255 |             g_loss, g_losses_latent = compute_g_loss(
256 |                 self.model, self.args.g_loss, x_real, y_org, y_trg, z_trgs=[z_trg, z_trg2], use_adv_cls=use_adv_cls)
257 |             g_loss, g_losses_ref = compute_g_loss(
258 |                 self.model, self.args.g_loss, x_real, y_org, y_trg, x_refs=[x_ref, x_ref2], use_adv_cls=use_adv_cls)
259 | 
260 |             for key in d_losses_latent:
261 |                 eval_losses["eval/%s" % key].append(d_losses_latent[key])
262 |             for key in g_losses_latent:
263 |                 eval_losses["eval/%s" % key].append(g_losses_latent[key])
264 | 
265 | #             if eval_steps_per_epoch % 10 == 0:
266 | #                 # generate x_fake
267 | #                 s_trg = self.model_ema.style_encoder(x_ref, y_trg)
268 | #                 F0 = self.model.f0_model.get_feature_GAN(x_real)
269 | #                 x_fake = self.model_ema.generator(x_real, s_trg, masks=None, F0=F0)
270 | #                 # generate x_recon
271 | #                 s_real = self.model_ema.style_encoder(x_real, y_org)
272 | #                 F0_fake = self.model.f0_model.get_feature_GAN(x_fake)
273 | #                 x_recon = self.model_ema.generator(x_fake, s_real, masks=None, F0=F0_fake)
274 |                 
275 | #                 eval_images['eval/image'].append(
276 | #                     ([x_real[0, 0].cpu().numpy(),
277 | #                     x_fake[0, 0].cpu().numpy(),
278 | #                     x_recon[0, 0].cpu().numpy()]))
279 | 
280 |         eval_losses = {key: np.mean(value) for key, value in eval_losses.items()}
281 |         eval_losses.update(eval_images)
282 |         return eval_losses
283 | 


--------------------------------------------------------------------------------
/transforms.py:
--------------------------------------------------------------------------------
  1 | # -*- coding: utf-8 -*-
  2 | 
  3 | import numpy as np
  4 | import torch
  5 | from torch import nn
  6 | import torch.nn.functional as F
  7 | import torchaudio
  8 | import torchaudio.functional as audio_F
  9 | import random
 10 | 
 11 | ## 1. RandomTimeStrech
 12 | 
 13 | class TimeStrech(nn.Module):
 14 |     def __init__(self, scale):
 15 |         super(TimeStrech, self).__init__()
 16 |         self.scale = scale
 17 | 
 18 |     def forward(self, x):
 19 |         mel_size = x.size(-1)
 20 |         
 21 |         x = F.interpolate(x, scale_factor=(1, self.scale), align_corners=False,
 22 |                           recompute_scale_factor=True, mode='bilinear').squeeze()
 23 |         
 24 |         if x.size(-1) < mel_size:
 25 |             noise_length = (mel_size - x.size(-1))
 26 |             random_pos = random.randint(0, x.size(-1)) - noise_length
 27 |             if random_pos < 0:
 28 |                 random_pos = 0
 29 |             noise = x[..., random_pos:random_pos + noise_length]
 30 |             x = torch.cat([x, noise], dim=-1)
 31 |         else:
 32 |             x = x[..., :mel_size]
 33 |         
 34 |         return x.unsqueeze(1)
 35 | 
 36 | ## 2. PitchShift
 37 | class PitchShift(nn.Module):
 38 |     def __init__(self, shift):
 39 |         super(PitchShift, self).__init__()
 40 |         self.shift = shift
 41 | 
 42 |     def forward(self, x):
 43 |         if len(x.shape) == 2:
 44 |             x = x.unsqueeze(0)
 45 |         x = x.squeeze()
 46 |         mel_size = x.shape[1]
 47 |         shift_scale = (mel_size + self.shift) / mel_size
 48 |         x = F.interpolate(x.unsqueeze(1), scale_factor=(shift_scale, 1.), align_corners=False,
 49 |                           recompute_scale_factor=True, mode='bilinear').squeeze(1)
 50 | 
 51 |         x = x[:, :mel_size]
 52 |         if x.size(1) < mel_size:
 53 |             pad_size = mel_size - x.size(1)
 54 |             x = torch.cat([x, torch.zeros(x.size(0), pad_size, x.size(2)).to(x.device)], dim=1)
 55 |         x = x.squeeze()
 56 |         return x.unsqueeze(1)
 57 | 
 58 | ## 3. ShiftBias
 59 | class ShiftBias(nn.Module):
 60 |     def __init__(self, bias):
 61 |         super(ShiftBias, self).__init__()
 62 |         self.bias = bias
 63 | 
 64 |     def forward(self, x):
 65 |         return x + self.bias
 66 | 
 67 | ## 4. Scaling
 68 | class SpectScaling(nn.Module):
 69 |     def __init__(self, scale):
 70 |         super(SpectScaling, self).__init__()
 71 |         self.scale = scale
 72 | 
 73 |     def forward(self, x):
 74 |         return x * self.scale
 75 | 
 76 | ## 5. Time Flip
 77 | class TimeFlip(nn.Module):
 78 |     def __init__(self, length):
 79 |         super(TimeFlip, self).__init__()
 80 |         self.length = round(length)
 81 | 
 82 |     def forward(self, x):
 83 |         if self.length > 1:
 84 |           start = np.random.randint(0, x.shape[-1] - self.length)
 85 |           x_ret = x.clone()
 86 |           x_ret[..., start:start + self.length] = torch.flip(x[..., start:start + self.length], dims=[-1])
 87 |           x = x_ret
 88 |         return x
 89 | 
 90 | class PhaseShuffle2d(nn.Module):
 91 |     def __init__(self, n=2):
 92 |         super(PhaseShuffle2d, self).__init__()
 93 |         self.n = n
 94 |         self.random = random.Random(1)
 95 | 
 96 |     def forward(self, x, move=None):
 97 |         # x.size = (B, C, M, L)
 98 |         if move is None:
 99 |             move = self.random.randint(-self.n, self.n)
100 | 
101 |         if move == 0:
102 |             return x
103 |         else:
104 |             left = x[:, :, :, :move]
105 |             right = x[:, :, :, move:]
106 |             shuffled = torch.cat([right, left], dim=3)
107 |             
108 |         return shuffled
109 | 
110 | def build_transforms():
111 |     transforms = [
112 |         lambda M: TimeStrech(1+ (np.random.random()-0.5)*M*0.2),
113 |         lambda M: SpectScaling(1 + (np.random.random()-1)*M*0.1),
114 |         lambda M: PhaseShuffle2d(192),
115 |     ]
116 |     N, M = len(transforms), np.random.random()
117 |     composed = nn.Sequential(
118 |         *[trans(M) for trans in np.random.choice(transforms, N)]
119 |     )
120 |     return composed
121 | 


--------------------------------------------------------------------------------