├── AISHELL-3
├── ReadMe.txt
└── spk-info.txt
├── LICENSE
├── README.md
├── aishell_data_clean.ipynb
├── asr.py
├── asr_v2.py
├── asr_v3.py
├── asr_web.py
├── config.py
├── infer_server.py
├── requirements.txt
├── static
├── index.css
├── record.js
└── record.png
├── templates
└── index.html
├── train.py
├── train_v2.py
├── train_v3.py
├── train_v4.py
└── utils
├── binary.py
├── callback.py
├── data_utils.py
├── model_utils.py
├── reader.py
└── utils.py
/AISHELL-3/ReadMe.txt:
--------------------------------------------------------------------------------
1 |
2 | AISHELL-3
3 |
4 | 北京希尔贝壳科技有限公司
5 | Beijing Shell Shell Technology Co.,Ltd
6 | 10/18/2020
7 |
8 | 1. AISHELL-3 Speech Data
9 |
10 | - Sampling Rate : 44.1kHz
11 | - Sample Format : 16bit
12 | - Environment : Quiet indoor
13 | - Speech Data Type : PCM
14 | - Channel Number : 1
15 | - Recording Equipment : High fidelity microphone
16 | - Sentences : 88035 utterances
17 | - Speaker : 218 speakers (43 male and 175 female)
18 |
19 |
20 | 2. Data Structure
21 | │ README.txt （readme）
22 | │ ChangeLog （Change Information）
23 | │ phone_set.txt （phone Information）
24 | │ spk_info.txt （Speaker Information）
25 | └─ test （Test Data File）
26 | └─ train （Train Data File）
27 | │├─content.txt （Transcript Content）
28 | │├─prosody_label_train-set.txt （Prosody Lable）
29 | │├─wav （Audio Data File）
30 | │├─SSB005 （Speaker ID File）
31 | ││ ││ ID2166W0001.wav （Audio）
32 |
33 | 4. System
34 | AISHELL-3 is a large-scale and high-fidelity multi-speaker Mandarin speech corpus which could be used to train multi-speaker Text-to-Speech (TTS) systems.
35 | You can download data set from: http://www.aishelltech.com/aishell_3.
36 | The baseline system code and generated samples are available online form: https://sos1sos2sixteen.github.io/aishell3/.
37 |
--------------------------------------------------------------------------------
/AISHELL-3/spk-info.txt:
--------------------------------------------------------------------------------
1 | # voice-file name; age group; gender; accent
2 | # In years, A:< 14, B:14 - 25, C:26 - 40, D:> 41.
3 |
4 | SSB1837 B female north
5 | SSB0578 B female north
6 | SSB1216 B female north
7 | SSB1161 B female north
8 | SSB0016 B female north
9 | SSB1365 B male north
10 | SSB1759 B female north
11 | SSB0588 B female north
12 | SSB0534 C female north
13 | SSB0380 B female north
14 | SSB0273 B male north
15 | SSB1863 B male north
16 | SSB1125 B female south
17 | SSB1872 B female north
18 | SSB0993 C female north
19 | SSB0666 C female north
20 | SSB0668 B female north
21 | SSB0395 B female north
22 | SSB1831 B male south
23 | SSB1064 B female north
24 | SSB1219 B female north
25 | SSB1322 B female north
26 | SSB1002 B female north
27 | SSB1878 B female north
28 | SSB0415 B female north
29 | SSB1781 C female north
30 | SSB0631 B male south
31 | SSB0686 B female north
32 | SSB1328 B male north
33 | SSB0366 B female south
34 | SSB1100 C male south
35 | SSB0241 B male north
36 | SSB0966 B male north
37 | SSB1340 B female north
38 | SSB0762 B female north
39 | SSB0073 B male north
40 | SSB0632 B female south
41 | SSB0915 B female south
42 | SSB0748 B female north
43 | SSB1956 B female north
44 | SSB1056 B female south
45 | SSB0716 B female south
46 | SSB0629 B male north
47 | SSB1806 B female north
48 | SSB0599 C female north
49 | SSB0720 B female north
50 | SSB0385 B female south
51 | SSB0809 B female north
52 | SSB0342 B female others
53 | SSB0760 B female north
54 | SSB1253 B female south
55 | SSB1575 B female south
56 | SSB0863 B male north
57 | SSB1110 D female north
58 | SSB0200 B female north
59 | SSB1215 B female north
60 | SSB0375 B male south
61 | SSB1828 C female north
62 | SSB0737 D female north
63 | SSB0341 C female north
64 | SSB0009 B female south
65 | SSB0309 D female north
66 | SSB1055 C female north
67 | SSB1448 B male north
68 | SSB1176 B female north
69 | SSB1001 B female north
70 | SSB0193 B female south
71 | SSB0710 C male north
72 | SSB0427 B female north
73 | SSB0338 B female north
74 | SSB1131 B female south
75 | SSB1108 B female north
76 | SSB0149 B female south
77 | SSB0736 B male south
78 | SSB1555 B female south
79 | SSB0614 C female north
80 | SSB1072 B female south
81 | SSB1728 B female north
82 | SSB1382 B female north
83 | SSB0851 B female north
84 | SSB1585 B female south
85 | SSB1891 C female north
86 | SSB1393 B female north
87 | SSB1274 B female north
88 | SSB1204 B female north
89 | SSB1452 B female north
90 | SSB0570 B female north
91 | SSB0780 B female north
92 | SSB1593 B female south
93 | SSB0913 B female north
94 | SSB1302 B female north
95 | SSB0323 B female north
96 | SSB1135 B female north
97 | SSB0382 B female north
98 | SSB0887 B male north
99 | SSB1625 C female south
100 | SSB1366 C female north
101 | SSB0693 B female south
102 | SSB0594 B female north
103 | SSB1686 B female north
104 | SSB0012 B female north
105 | SSB0139 B male south
106 | SSB0751 B female north
107 | SSB0606 D female north
108 | SSB1341 B female south
109 | SSB0145 B female north
110 | SSB1136 B male south
111 | SSB0339 B female north
112 | SSB0482 B female north
113 | SSB0502 B female north
114 | SSB1650 B female north
115 | SSB0817 B male north
116 | SSB0261 C male north
117 | SSB0316 B male north
118 | SSB0033 B female south
119 | SSB0723 B female north
120 | SSB1008 B female north
121 | SSB0700 B female north
122 | SSB1457 B female north
123 | SSB0601 B female north
124 | SSB1809 B female north
125 | SSB1739 B female north
126 | SSB0407 C male north
127 | SSB0426 B female south
128 | SSB0470 B female north
129 | SSB0935 B female north
130 | SSB0822 B female north
131 | SSB0746 B female north
132 | SSB0758 B female north
133 | SSB1221 B female north
134 | SSB0038 B female north
135 | SSB1624 B male south
136 | SSB0133 B female south
137 | SSB0778 B female north
138 | SSB0702 B female south
139 | SSB1383 B male south
140 | SSB1563 B female south
141 | SSB1670 B female north
142 | SSB1096 B female south
143 | SSB0299 B female south
144 | SSB1711 B female north
145 | SSB1810 B female north
146 | SSB1115 B female south
147 | SSB1684 B male north
148 | SSB1402 C female north
149 | SSB1918 B female north
150 | SSB0246 B female south
151 | SSB1607 B female north
152 | SSB1437 C female north
153 | SSB0011 C female north
154 | SSB0288 C female north
155 | SSB0539 C female north
156 | SSB0394 B male north
157 | SSB0379 B female south
158 | SSB1187 B male north
159 | SSB0671 C female north
160 | SSB0544 B male north
161 | SSB0005 B female north
162 | SSB1846 B female south
163 | SSB0122 B female north
164 | SSB1630 C male north
165 | SSB1399 B female north
166 | SSB1567 B female others
167 | SSB0267 C female north
168 | SSB1392 B female north
169 | SSB0287 B female north
170 | SSB0717 B female south
171 | SSB0018 B female south
172 | SSB0315 B female south
173 | SSB1126 A female north
174 | SSB1320 D female north
175 | SSB0919 B female north
176 | SSB0623 B male south
177 | SSB0871 C female north
178 | SSB0786 B female north
179 | SSB1024 B female north
180 | SSB1935 B male north
181 | SSB0794 B female north
182 | SSB1832 B female north
183 | SSB1000 C female north
184 | SSB1431 C female north
185 | SSB0535 B male north
186 | SSB1782 C female north
187 | SSB0354 D female north
188 | SSB0393 A female north
189 | SSB0057 B female north
190 | SSB1385 B female north
191 | SSB1050 D female north
192 | SSB0435 B female north
193 | SSB1197 B female north
194 | SSB1939 B female north
195 | SSB1239 B male north
196 | SSB0434 D male north
197 | SSB1020 B male south
198 | SSB1091 B female north
199 | SSB1745 B male north
200 | SSB0565 B female north
201 | SSB0711 B female south
202 | SSB0603 B male north
203 | SSB0749 B female south
204 | SSB0997 B female north
205 | SSB0590 C male north
206 | SSB0784 C male north
207 | SSB1699 C female north
208 | SSB0607 C female north
209 | SSB1902 B female north
210 | SSB0043 B female north
211 | SSB1138 B female north
212 | SSB1408 B male north
213 | SSB1218 B female south
214 | SSB1377 B female north
215 | SSB0609 C male north
216 | SSB0987 B female north
217 | SSB0080 B female south
218 | SSB0307 B female north
219 | SSB0197 C female south
220 | SSB1203 C female north
221 | SSB0112 B female south
222 |
--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
1 | MIT License
2 |
3 | Copyright (c) 2024 Thirteen
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 | # asr_AISHELL-3
2 | 使用AISHELL-3 数据集 训练语音识别模型
3 |
4 | ## 使用方法
5 | 创建虚拟环境
6 | ```
7 | conda create -n asr python=3.10
8 | ```
9 | 配置环境
10 | ```
11 | activate asr
12 | pip install -r requirements.txt
13 | ```
14 | 运行
15 | ```
16 | python train.py
17 | ```
18 | 如果已经运行且已经有音频特征文件 features.pkl
19 | 可直接运行 __trian_v2.py__
20 | ```
21 | python train_v2.py
22 | ```
23 | * 在train_v4中增加tensorboard
24 | 可查看训练日志
25 | ## 使用web进行语音识别
26 | 运行
27 | ```
28 | python asr_web.py
29 | ```
30 | * 可进行读取录音
31 | * 本地录制并上传进行识别
32 | * 预览
33 | 
34 |
35 | ## 查看训练日志
36 | 输入命令
37 | ```
38 | tensorboard --logdir= log_path
39 | ```
40 | ## loss曲线：
41 | __epochs=25__
42 | 
43 |
44 | ## librosa版本问题
45 | 这里使用的 *librosa==0.7.2*
46 | 可能会出现
47 | 
48 | **参考这份博客：**
49 |
50 | [解决不联网环境pip安装librosa、numba、llvmlite报错和版本兼容问题](https://blog.csdn.net/qq_39691492/article/details/130829401)
51 |
52 | *修改如下：*
53 |
54 | 
55 |
56 | ## api的部分参考
57 | [Whisper-Finetune](https://github.com/yeyupiaoling/Whisper-Finetune)
58 |
59 | 将原来微调的whisper模型换成这里训练的asr模型
60 |
--------------------------------------------------------------------------------
/aishell_data_clean.ipynb:
--------------------------------------------------------------------------------
1 | {
2 | "cells": [
3 | {
4 | "cell_type": "code",
5 | "execution_count": 14,
6 | "metadata": {},
7 | "outputs": [],
8 | "source": [
9 | "import os\n",
10 | "path = 'AISHELL-3/train/wav/'\n",
11 | "files = os.listdir(path)"
12 | ]
13 | },
14 | {
15 | "cell_type": "code",
16 | "execution_count": 48,
17 | "metadata": {},
18 | "outputs": [],
19 | "source": [
20 | "# 音频文件\n",
21 | "wav_file=[]\n",
22 | "# 音频文件的相对路径\n",
23 | "wav_file_path=[]\n",
24 | "\n",
25 | "for i in files:\n",
26 | " temp_files = os.listdir(path+i)\n",
27 | " for j in temp_files:\n",
28 | " wav_file.append(j)\n",
29 | " wav_file_path.append(path+i+'/'+j)\n",
30 | "\n"
31 | ]
32 | },
33 | {
34 | "cell_type": "markdown",
35 | "metadata": {},
36 | "source": [
37 | "* 两种导入方式"
38 | ]
39 | },
40 | {
41 | "cell_type": "code",
42 | "execution_count": 54,
43 | "metadata": {},
44 | "outputs": [],
45 | "source": [
46 | "# with open(\"AISHELL-3/train/content.txt\", \"r\", encoding='utf-8') as f: # 打开文件\n",
47 | "# data = f.read() # 读取文件\n",
48 | " # print(data)"
49 | ]
50 | },
51 | {
52 | "cell_type": "code",
53 | "execution_count": 60,
54 | "metadata": {},
55 | "outputs": [],
56 | "source": [
57 | "f=open(\"AISHELL-3/train/content.txt\", \"r\", encoding='utf-8')\n",
58 | "txt=[]\n",
59 | "for line in f:\n",
60 | " txt.append(line.strip())\n"
61 | ]
62 | },
63 | {
64 | "cell_type": "code",
65 | "execution_count": 66,
66 | "metadata": {},
67 | "outputs": [],
68 | "source": [
69 | "# 音频文件名字\n",
70 | "txt_filename = []\n",
71 | "# 对应文件内容\n",
72 | "txt_wav = []\n",
73 | "\n",
74 | "for i in txt:\n",
75 | " temp_txt_filename,temp_txt_wav = i.split('\\t')\n",
76 | " txt_filename.append(temp_txt_filename)\n",
77 | " txt_wav.append(temp_txt_wav)"
78 | ]
79 | },
80 | {
81 | "cell_type": "code",
82 | "execution_count": 76,
83 | "metadata": {},
84 | "outputs": [],
85 | "source": [
86 | "# 字典 字母+数字\n",
87 | "en_num_all = []\n",
88 | "\n",
89 | "for letter in 'abcdefghijklmnopqrstuvwxyz':\n",
90 | " en_num_all.extend(letter)\n",
91 | "\n",
92 | "for number in range(10): \n",
93 | " en_num_all.extend(str(number))\n",
94 | " \n",
95 | "en_num_all.extend(' ')"
96 | ]
97 | },
98 | {
99 | "cell_type": "code",
100 | "execution_count": 109,
101 | "metadata": {},
102 | "outputs": [],
103 | "source": [
104 | "# 音频文字\n",
105 | "texts = []\n",
106 | "\n",
107 | "for i in txt_wav:\n",
108 | " temp = ''\n",
109 | " for j in i:\n",
110 | " if j in en_num_all:\n",
111 | " continue\n",
112 | " else:\n",
113 | " # print(j)\n",
114 | " temp = temp+j\n",
115 | " texts.append(temp)\n",
116 | " "
117 | ]
118 | },
119 | {
120 | "cell_type": "markdown",
121 | "metadata": {},
122 | "source": [
123 | "#### texts 与 wav_file_path 对应"
124 | ]
125 | },
126 | {
127 | "cell_type": "code",
128 | "execution_count": 110,
129 | "metadata": {},
130 | "outputs": [
131 | {
132 | "data": {
133 | "text/plain": [
134 | "'一百三十三万三千五百六十五'"
135 | ]
136 | },
137 | "execution_count": 110,
138 | "metadata": {},
139 | "output_type": "execute_result"
140 | }
141 | ],
142 | "source": [
143 | "texts[-1]"
144 | ]
145 | },
146 | {
147 | "cell_type": "code",
148 | "execution_count": 107,
149 | "metadata": {},
150 | "outputs": [
151 | {
152 | "data": {
153 | "text/plain": [
154 | "'AISHELL-3/train/wav/SSB1956/SSB19560481.wav'"
155 | ]
156 | },
157 | "execution_count": 107,
158 | "metadata": {},
159 | "output_type": "execute_result"
160 | }
161 | ],
162 | "source": [
163 | "wav_file_path[-1]"
164 | ]
165 | }
166 | ],
167 | "metadata": {
168 | "kernelspec": {
169 | "display_name": "Python 3",
170 | "language": "python",
171 | "name": "python3"
172 | },
173 | "language_info": {
174 | "codemirror_mode": {
175 | "name": "ipython",
176 | "version": 3
177 | },
178 | "file_extension": ".py",
179 | "mimetype": "text/x-python",
180 | "name": "python",
181 | "nbconvert_exporter": "python",
182 | "pygments_lexer": "ipython3",
183 | "version": "3.10.2"
184 | }
185 | },
186 | "nbformat": 4,
187 | "nbformat_minor": 2
188 | }
189 |
--------------------------------------------------------------------------------
/asr.py:
--------------------------------------------------------------------------------
1 | from keras.models import load_model
2 | from keras import backend as K
3 | import numpy as np
4 | import librosa
5 | from python_speech_features import mfcc
6 | import speech_recognition as sr
7 | import pickle
8 | import glob
9 | import config
10 | import wave
11 | import os
12 |
13 | def save_as_wav(audio, output_file_path):
14 | with wave.open(output_file_path, 'wb') as wav_file:
15 | wav_file.setnchannels(1) # 单声道
16 | wav_file.setsampwidth(2) # 16位PCM编码
17 | wav_file.setframerate(44100) # 采样率为44.1kHz
18 | wav_file.writeframes(audio.frame_data)
19 |
20 | def input_audio():
21 | r = sr.Recognizer()
22 | with sr.Microphone() as source:
23 | print("请说...")
24 | r.pause_threshold = 1
25 | audio = r.listen(source)
26 | output_file_path = "temp_file.wav"
27 | save_as_wav(audio, output_file_path)
28 | wavs = glob.glob('temp_file.wav')
29 | return wavs
30 |
31 | def load_file():
32 | with open(config.pkl_path, 'rb') as fr:
33 | [char2id, id2char, mfcc_mean, mfcc_std] = pickle.load(fr)
34 | model = load_model(config.model_path)
35 | return char2id, id2char, mfcc_mean, mfcc_std, model
36 |
37 | def set_data(wavs, mfcc_mean, mfcc_std):
38 | mfcc_dim = config.mfcc_dim
39 | index = np.random.randint(len(wavs))
40 | audio, sr = librosa.load(wavs[index])
41 | energy = librosa.feature.rms(audio)
42 | frames = np.nonzero(energy >= np.max(energy) / 5)
43 | indices = librosa.core.frames_to_samples(frames)[1]
44 | audio = audio[indices[0]:indices[-1]] if indices.size else audio[0:0]
45 | X_data = mfcc(audio, sr, numcep=mfcc_dim, nfft=551)
46 | X_data = (X_data - mfcc_mean) / (mfcc_std + 1e-14)
47 | print(X_data.shape)
48 | return X_data
49 |
50 |
51 | def wav_pred(model,X_data,id2char):
52 | pred = model.predict(np.expand_dims(X_data, axis=0))
53 | pred_ids = K.eval(K.ctc_decode(pred, [X_data.shape[0]], greedy=False, beam_width=10, top_paths=1)[0][0])
54 | pred_ids = pred_ids.flatten().tolist()
55 | words=''
56 | judge=0
57 | for i in pred_ids:
58 | if i != -1:
59 | judge=1
60 | words=words+id2char[i]
61 | if judge==1:
62 | print(words)
63 | else:
64 | print("未检测到")
65 |
66 | def run():
67 | wavs = input_audio()
68 | char2id, id2char, mfcc_mean, mfcc_std, model = load_file()
69 | X_data = set_data(wavs, mfcc_mean, mfcc_std)
70 | wav_pred(model,X_data,id2char)
71 | os.remove("temp_file.wav")
72 |
73 |
74 | if __name__ == '__main__' :
75 | run()
76 |
77 |
--------------------------------------------------------------------------------
/asr_v2.py:
--------------------------------------------------------------------------------
1 | from keras.models import load_model
2 | from keras import backend as K
3 | import numpy as np
4 | import librosa
5 | from python_speech_features import mfcc
6 | import speech_recognition as sr
7 | import pickle
8 | import glob
9 | import config
10 | import wave
11 | import os
12 |
13 | def save_as_wav(audio, output_file_path):
14 | with wave.open(output_file_path, 'wb') as wav_file:
15 | wav_file.setnchannels(1) # 单声道
16 | wav_file.setsampwidth(2) # 16位PCM编码
17 | wav_file.setframerate(44100) # 采样率为44.1kHz
18 | wav_file.writeframes(audio.frame_data)
19 |
20 | def input_audio():
21 | r = sr.Recognizer()
22 | with sr.Microphone() as source:
23 | print("请说...")
24 | r.pause_threshold = 1
25 | audio = r.listen(source)
26 | output_file_path = "temp_file.wav"
27 | save_as_wav(audio, output_file_path)
28 | wavs = glob.glob('temp_file.wav')
29 | os.remove("temp_file.wav")
30 | return wavs
31 |
32 | def out_load_audio():
33 | path = config.audio_path
34 | wavs = glob.glob(path)
35 | return wavs
36 |
37 | def load_file():
38 | with open(config.pkl_path, 'rb') as fr:
39 | [char2id, id2char, mfcc_mean, mfcc_std] = pickle.load(fr)
40 | model = load_model(config.model_path)
41 | return char2id, id2char, mfcc_mean, mfcc_std, model
42 |
43 | def set_data(wavs, mfcc_mean, mfcc_std):
44 | mfcc_dim = config.mfcc_dim
45 | index = np.random.randint(len(wavs))
46 | audio, sr = librosa.load(wavs[index])
47 | energy = librosa.feature.rms(audio)
48 | frames = np.nonzero(energy >= np.max(energy) / 5)
49 | indices = librosa.core.frames_to_samples(frames)[1]
50 | audio = audio[indices[0]:indices[-1]] if indices.size else audio[0:0]
51 | X_data = mfcc(audio, sr, numcep=mfcc_dim, nfft=551)
52 | X_data = (X_data - mfcc_mean) / (mfcc_std + 1e-14)
53 | # print(X_data.shape)
54 | return X_data
55 |
56 |
57 | def wav_pred(model,X_data,id2char):
58 | pred = model.predict(np.expand_dims(X_data, axis=0))
59 | pred_ids = K.eval(K.ctc_decode(pred, [X_data.shape[0]], greedy=False, beam_width=10, top_paths=1)[0][0])
60 | pred_ids = pred_ids.flatten().tolist()
61 | words=''
62 | judge=0
63 | for i in pred_ids:
64 | if i != -1:
65 | judge=1
66 | words=words+id2char[i]
67 | if judge==1:
68 | print(words)
69 | else:
70 | print("未检测到")
71 |
72 | def run():
73 | # wavs = input_audio()
74 | wavs = out_load_audio()
75 | char2id, id2char, mfcc_mean, mfcc_std, model = load_file()
76 | X_data = set_data(wavs, mfcc_mean, mfcc_std)
77 | wav_pred(model,X_data,id2char)
78 |
79 |
80 |
81 | if __name__ == '__main__' :
82 | run()
83 |
84 |
--------------------------------------------------------------------------------
/asr_v3.py:
--------------------------------------------------------------------------------
1 | from keras.models import load_model
2 | from keras import backend as K
3 | import numpy as np
4 | import librosa
5 | from python_speech_features import mfcc
6 | import speech_recognition as sr
7 | import pickle
8 | import glob
9 | import config
10 | import wave
11 | import os
12 | import pyaudio
13 | from tqdm import tqdm
14 |
15 | class set_audio():
16 | CHUNK = 1024 # 每个缓冲区的帧数
17 | FORMAT = pyaudio.paInt16 # 采样位数
18 | CHANNELS = 1 # 单声道
19 | RATE = 44100 # 采样频率
20 |
21 | # 可设置录制时间
22 | def record_audio(record_second):
23 | """ 录音功能 """
24 | p = pyaudio.PyAudio() # 实例化对象
25 | stream = p.open(format=set_audio.FORMAT,
26 | channels=set_audio.CHANNELS,
27 | rate=set_audio.RATE,
28 | input=True,
29 | frames_per_buffer=set_audio.CHUNK) # 打开流，传入响应参数
30 |
31 | wf = wave.open('temp_file.wav', 'wb') # 打开 wav 文件。
32 | wf.setnchannels(set_audio.CHANNELS) # 声道设置
33 | wf.setsampwidth(p.get_sample_size(set_audio.FORMAT)) # 采样位数设置
34 | wf.setframerate(set_audio.RATE) # 采样频率设置
35 |
36 | for _ in tqdm(range(0, int(set_audio.RATE * record_second / set_audio.CHUNK))):
37 | data = stream.read(set_audio.CHUNK)
38 | wf.writeframes(data) # 写入数据
39 | stream.stop_stream() # 关闭流
40 | stream.close()
41 | p.terminate()
42 | wf.close()
43 |
44 | wavs = glob.glob('temp_file.wav')
45 |
46 | # os.remove("temp_file.wav")
47 |
48 | return wavs
49 |
50 |
51 |
52 | def save_as_wav(audio, output_file_path):
53 | with wave.open(output_file_path, 'wb') as wav_file:
54 | wav_file.setnchannels(1) # 单声道
55 | wav_file.setsampwidth(2) # 16位PCM编码
56 | wav_file.setframerate(44100) # 采样率为44.1kHz
57 | wav_file.writeframes(audio.frame_data)
58 |
59 | # 录音自动停止
60 | def input_audio():
61 | r = sr.Recognizer()
62 | with sr.Microphone() as source:
63 | print("请说...")
64 | r.pause_threshold = 1
65 | audio = r.listen(source)
66 | output_file_path = "temp_file.wav"
67 | save_as_wav(audio, output_file_path)
68 | wavs = glob.glob('temp_file.wav')
69 |
70 | return wavs
71 |
72 | def load_file():
73 | with open(config.pkl_path, 'rb') as fr:
74 | [char2id, id2char, mfcc_mean, mfcc_std] = pickle.load(fr)
75 | model = load_model(config.model_path)
76 | return char2id, id2char, mfcc_mean, mfcc_std, model
77 |
78 | def set_data(wavs, mfcc_mean, mfcc_std):
79 | mfcc_dim = config.mfcc_dim
80 | index = np.random.randint(len(wavs))
81 | audio, sr = librosa.load(wavs[index])
82 | energy = librosa.feature.rms(audio)
83 | frames = np.nonzero(energy >= np.max(energy) / 5)
84 | indices = librosa.core.frames_to_samples(frames)[1]
85 | audio = audio[indices[0]:indices[-1]] if indices.size else audio[0:0]
86 | X_data = mfcc(audio, sr, numcep=mfcc_dim, nfft=551)
87 | X_data = (X_data - mfcc_mean) / (mfcc_std + 1e-14)
88 | # print(X_data.shape)
89 | return X_data
90 |
91 |
92 | def wav_pred(model,X_data,id2char):
93 | pred = model.predict(np.expand_dims(X_data, axis=0))
94 | pred_ids = K.eval(K.ctc_decode(pred, [X_data.shape[0]], greedy=False, beam_width=10, top_paths=1)[0][0])
95 | pred_ids = pred_ids.flatten().tolist()
96 | words=''
97 | judge=0
98 | for i in pred_ids:
99 | if i != -1:
100 | judge=1
101 | words=words+id2char[i]
102 | if judge==1:
103 | print(words)
104 | else:
105 | print("未检测到")
106 |
107 | def run():
108 | # 自动停止录音
109 | # wavs = input_audio()
110 | # 设置录制时间
111 | wavs = record_audio(record_second=5)
112 | char2id, id2char, mfcc_mean, mfcc_std, model = load_file()
113 | X_data = set_data(wavs, mfcc_mean, mfcc_std)
114 | wav_pred(model,X_data,id2char)
115 | os.remove("temp_file.wav")
116 |
117 |
118 | if __name__ == '__main__' :
119 | run()
120 |
121 |
--------------------------------------------------------------------------------
/asr_web.py:
--------------------------------------------------------------------------------
1 | import argparse
2 | import functools
3 | import os
4 |
5 | import torch
6 | import uvicorn
7 | from fastapi import FastAPI, File, Body, UploadFile, Request
8 | from starlette.staticfiles import StaticFiles
9 | from starlette.templating import Jinja2Templates
10 | from utils.utils import add_arguments, print_arguments
11 |
12 | from keras.models import load_model
13 | from keras import backend as K
14 | import numpy as np
15 | import librosa
16 | from python_speech_features import mfcc
17 | import speech_recognition as sr
18 | import pickle
19 | import config
20 | import wave
21 | import io
22 |
23 | from pydub import AudioSegment
24 | from io import BytesIO
25 | import librosa
26 | import numpy as np
27 | from python_speech_features import mfcc
28 |
29 | os.environ['KMP_DUPLICATE_LIB_OK'] = 'True'
30 | parser = argparse.ArgumentParser(description=__doc__)
31 | add_arg = functools.partial(add_arguments, argparser=parser)
32 | add_arg("host", type=str, default="0.0.0.0", help="监听主机的IP地址")
33 | add_arg("port", type=int, default=5000, help="服务所使用的端口号")
34 | # add_arg("model_path", type=str, default="models/whisper-tiny-finetune/", help="合并模型的路径，或者是huggingface上模型的名称")
35 | # add_arg("model_path", type=str, default="models/tiny-finetune/", help="合并模型的路径，或者是huggingface上模型的名称")
36 | add_arg("use_gpu", type=bool, default=True, help="是否使用gpu进行预测")
37 | add_arg("num_beams", type=int, default=1, help="解码搜索大小")
38 | add_arg("batch_size", type=int, default=16, help="预测batch_size大小")
39 | add_arg("use_compile", type=bool, default=False, help="是否使用Pytorch2.0的编译器")
40 | add_arg("assistant_model_path", type=str, default=None, help="助手模型，可以提高推理速度，例如openai/whisper-tiny")
41 | add_arg("local_files_only", type=bool, default=True, help="是否只在本地加载模型，不尝试下载")
42 | add_arg("use_flash_attention_2", type=bool, default=False, help="是否使用FlashAttention2加速")
43 | add_arg("use_bettertransformer", type=bool, default=False, help="是否使用BetterTransformer加速")
44 | args = parser.parse_args()
45 | print_arguments(args)
46 |
47 | # 设置设备
48 | device = "cuda:0" if torch.cuda.is_available() and args.use_gpu else "cpu"
49 | torch_dtype = torch.float16 if torch.cuda.is_available() and args.use_gpu else torch.float32
50 |
51 |
52 | model = load_model(config.model_path)
53 |
54 | with open(config.pkl_path, 'rb') as fr:
55 | [char2id, id2char, mfcc_mean, mfcc_std] = pickle.load(fr)
56 |
57 |
58 | app = FastAPI(title="thirteen语音识别")
59 | app.mount('/static', StaticFiles(directory='static'), name='static')
60 | templates = Jinja2Templates(directory="templates")
61 | model_semaphore = None
62 |
63 |
64 | def release_model_semaphore():
65 | model_semaphore.release()
66 |
67 |
68 | def recognition(file: File,mfcc_mean, mfcc_std):
69 |
70 | X_data = extract_mfcc_features(file, mfcc_mean, mfcc_std)
71 | pred = model.predict(np.expand_dims(X_data, axis=0))
72 | pred_ids = K.eval(K.ctc_decode(pred, [X_data.shape[0]], greedy=False, beam_width=10, top_paths=1)[0][0])
73 | pred_ids = pred_ids.flatten().tolist()
74 | results = ''
75 | judge=0
76 | for i in pred_ids:
77 | if i != -1:
78 | judge=1
79 | results = results + id2char[i]
80 | if judge!=1:
81 | results = '未检测到'
82 |
83 | return results
84 |
85 |
86 | def extract_mfcc_features(audio_bytes, mfcc_mean, mfcc_std):
87 | # 使用pydub将bytes转换为WAV格式的AudioSegment（如果它不是WAV的话）
88 | # 注意：这里我们假设input_bytes是WAV或我们可以转换为WAV的格式
89 | # 如果input_bytes不是WAV且格式未知，你可能需要先检测它
90 | audio_segment = AudioSegment.from_file(BytesIO(audio_bytes), format="wav") # 如果已经是WAV，或者确定可以解析为WAV
91 |
92 | # 确保输出为WAV格式（如果之前不是的话，这一步其实是多余的，因为from_file已经处理了）
93 | # 但为了清晰起见，我们还是将其导出为WAV的bytes
94 | wav_bytes = BytesIO()
95 | audio_segment.export(wav_bytes, format="wav")
96 |
97 | # 重置BytesIO的指针到开头
98 | wav_bytes.seek(0)
99 |
100 | # 使用librosa加载WAV音频
101 | y, sr = librosa.load(wav_bytes)
102 |
103 | # 提取RMS能量
104 | energy = librosa.feature.rms(y=y)
105 |
106 | # 找到能量大于最大能量1/5的帧
107 | frames = np.nonzero(energy[0] >= np.max(energy[0]) / 5)
108 |
109 | # 将帧索引转换为样本索引
110 | if frames[0].size:
111 | indices = librosa.core.frames_to_samples(frames)[0]
112 | y = y[indices[0]:indices[-1]]
113 |
114 | # 提取MFCC特征
115 | mfcc_dim = 13 # 你可以根据需要修改MFCC的维度
116 | mfcc_features = mfcc(y, sr, numcep=mfcc_dim, nfft=551)
117 |
118 | # 这里假设你已经有了mfcc_mean和mfcc_std用于标准化（通常需要在训练阶段计算）
119 | # 如果没有，你可以跳过标准化步骤，或者计算它们
120 | mfcc_features = (mfcc_features - mfcc_mean) / (mfcc_std + 1e-14)
121 |
122 | return mfcc_features
123 |
124 |
125 | @app.post("/recognition")
126 | async def api_recognition(audio: UploadFile = File(..., description="音频文件")):
127 | # if language == "None": language = None
128 | data = await audio.read()
129 | with io.BytesIO(data) as bio:
130 | with wave.open(bio, 'rb') as wav_file:
131 | pass
132 | results = recognition(file= data, mfcc_mean= mfcc_mean, mfcc_std= mfcc_std)
133 | ret = {"results": results, "code": 0}
134 | return ret
135 |
136 |
137 | @app.get("/")
138 | async def index(request: Request):
139 | return templates.TemplateResponse("index.html", {"request": request, "id": id})
140 |
141 |
142 | if __name__ == '__main__':
143 | uvicorn.run(app, host=args.host, port=args.port)
144 |
--------------------------------------------------------------------------------
/config.py:
--------------------------------------------------------------------------------
1 | # 训练数据音频文件路径
2 | train_wav_data_path = 'AISHELL-3/train/wav/'
3 |
4 | # 训练数据内容文件路径
5 | train_texts_data_path = 'AISHELL-3/train/content.txt'
6 |
7 | # 测试数据音频文件路径
8 | test_wav_data_path = 'AISHELL-3/test/wav/'
9 |
10 | # 测试数据内容文件路径
11 | test_texts_data_path = 'AISHELL-3/test/content.txt'
12 |
13 | # 存放模型路径 /模型名字
14 | model_path = 'model/asr_AISHELL.h5'
15 |
16 | # 存放pkl路径 /pkl名字
17 | pkl_path = 'pkl_all/dictionary.pkl'
18 |
19 | # 存放labels路径
20 | labels_path = 'pkl_all/labels.pkl'
21 |
22 | # features.pkl路径
23 | features_path = 'pkl_all/features.pkl'
24 |
25 | # 外部导入音频路径
26 | audio_path = 'AISHELL-3/train/wav/SSB0005/SSB00050001.wav'
27 | # model_name
28 |
29 | batch_size = 16
30 |
31 | epochs = 25
32 |
33 | num_blocks = 3
34 |
35 | filters = 128
36 |
37 | mfcc_dim = 13
38 |
--------------------------------------------------------------------------------
/infer_server.py:
--------------------------------------------------------------------------------
1 | import argparse
2 | import functools
3 | import os
4 |
5 | import torch
6 | import uvicorn
7 | from fastapi import FastAPI, File, Body, UploadFile, Request
8 | from starlette.staticfiles import StaticFiles
9 | from starlette.templating import Jinja2Templates
10 | from utils.utils import add_arguments, print_arguments
11 |
12 | from keras.models import load_model
13 | from keras import backend as K
14 | import numpy as np
15 | import librosa
16 | from python_speech_features import mfcc
17 | import speech_recognition as sr
18 | import pickle
19 | import config
20 | import wave
21 | import io
22 |
23 | from pydub import AudioSegment
24 | from io import BytesIO
25 | import librosa
26 | import numpy as np
27 | from python_speech_features import mfcc
28 |
29 | os.environ['KMP_DUPLICATE_LIB_OK'] = 'True'
30 | parser = argparse.ArgumentParser(description=__doc__)
31 | add_arg = functools.partial(add_arguments, argparser=parser)
32 | add_arg("host", type=str, default="0.0.0.0", help="监听主机的IP地址")
33 | add_arg("port", type=int, default=5000, help="服务所使用的端口号")
34 | # add_arg("model_path", type=str, default="models/whisper-tiny-finetune/", help="合并模型的路径，或者是huggingface上模型的名称")
35 | # add_arg("model_path", type=str, default="models/tiny-finetune/", help="合并模型的路径，或者是huggingface上模型的名称")
36 | add_arg("use_gpu", type=bool, default=True, help="是否使用gpu进行预测")
37 | add_arg("num_beams", type=int, default=1, help="解码搜索大小")
38 | add_arg("batch_size", type=int, default=16, help="预测batch_size大小")
39 | add_arg("use_compile", type=bool, default=False, help="是否使用Pytorch2.0的编译器")
40 | add_arg("assistant_model_path", type=str, default=None, help="助手模型，可以提高推理速度，例如openai/whisper-tiny")
41 | add_arg("local_files_only", type=bool, default=True, help="是否只在本地加载模型，不尝试下载")
42 | add_arg("use_flash_attention_2", type=bool, default=False, help="是否使用FlashAttention2加速")
43 | add_arg("use_bettertransformer", type=bool, default=False, help="是否使用BetterTransformer加速")
44 | args = parser.parse_args()
45 | print_arguments(args)
46 |
47 | # 设置设备
48 | device = "cuda:0" if torch.cuda.is_available() and args.use_gpu else "cpu"
49 | torch_dtype = torch.float16 if torch.cuda.is_available() and args.use_gpu else torch.float32
50 |
51 |
52 | model = load_model(config.model_path)
53 |
54 | with open(config.pkl_path, 'rb') as fr:
55 | [char2id, id2char, mfcc_mean, mfcc_std] = pickle.load(fr)
56 |
57 |
58 | app = FastAPI(title="thirteen语音识别")
59 | app.mount('/static', StaticFiles(directory='static'), name='static')
60 | templates = Jinja2Templates(directory="templates")
61 | model_semaphore = None
62 |
63 |
64 | def release_model_semaphore():
65 | model_semaphore.release()
66 |
67 |
68 | def recognition(file: File,mfcc_mean, mfcc_std):
69 |
70 | X_data = extract_mfcc_features(file, mfcc_mean, mfcc_std)
71 | pred = model.predict(np.expand_dims(X_data, axis=0))
72 | pred_ids = K.eval(K.ctc_decode(pred, [X_data.shape[0]], greedy=False, beam_width=10, top_paths=1)[0][0])
73 | pred_ids = pred_ids.flatten().tolist()
74 | results = ''
75 | judge=0
76 | for i in pred_ids:
77 | if i != -1:
78 | judge=1
79 | results = results + id2char[i]
80 | if judge!=1:
81 | results = '未检测到'
82 |
83 | return results
84 |
85 |
86 | def extract_mfcc_features(audio_bytes, mfcc_mean, mfcc_std):
87 | # 使用pydub将bytes转换为WAV格式的AudioSegment（如果它不是WAV的话）
88 | # 注意：这里我们假设input_bytes是WAV或我们可以转换为WAV的格式
89 | # 如果input_bytes不是WAV且格式未知，你可能需要先检测它
90 | audio_segment = AudioSegment.from_file(BytesIO(audio_bytes), format="wav") # 如果已经是WAV，或者确定可以解析为WAV
91 |
92 | # 确保输出为WAV格式（如果之前不是的话，这一步其实是多余的，因为from_file已经处理了）
93 | # 但为了清晰起见，我们还是将其导出为WAV的bytes
94 | wav_bytes = BytesIO()
95 | audio_segment.export(wav_bytes, format="wav")
96 |
97 | # 重置BytesIO的指针到开头
98 | wav_bytes.seek(0)
99 |
100 | # 使用librosa加载WAV音频
101 | y, sr = librosa.load(wav_bytes)
102 |
103 | # 提取RMS能量
104 | energy = librosa.feature.rms(y=y)
105 |
106 | # 找到能量大于最大能量1/5的帧
107 | frames = np.nonzero(energy[0] >= np.max(energy[0]) / 5)
108 |
109 | # 将帧索引转换为样本索引
110 | if frames[0].size:
111 | indices = librosa.core.frames_to_samples(frames)[0]
112 | y = y[indices[0]:indices[-1]]
113 |
114 | # 提取MFCC特征
115 | mfcc_dim = 13 # 你可以根据需要修改MFCC的维度
116 | mfcc_features = mfcc(y, sr, numcep=mfcc_dim, nfft=551)
117 |
118 | # 这里假设你已经有了mfcc_mean和mfcc_std用于标准化（通常需要在训练阶段计算）
119 | # 如果没有，你可以跳过标准化步骤，或者计算它们
120 | mfcc_features = (mfcc_features - mfcc_mean) / (mfcc_std + 1e-14)
121 |
122 | return mfcc_features
123 |
124 |
125 | @app.post("/recognition")
126 | async def api_recognition(audio: UploadFile = File(..., description="音频文件")):
127 | # if language == "None": language = None
128 | data = await audio.read()
129 | with io.BytesIO(data) as bio:
130 | with wave.open(bio, 'rb') as wav_file:
131 | pass
132 | results = recognition(file= data, mfcc_mean= mfcc_mean, mfcc_std= mfcc_std)
133 | ret = {"results": results, "code": 0}
134 | return ret
135 |
136 |
137 | @app.get("/")
138 | async def index(request: Request):
139 | return templates.TemplateResponse("index.html", {"request": request, "id": id})
140 |
141 |
142 | if __name__ == '__main__':
143 | uvicorn.run(app, host=args.host, port=args.port)
144 |
--------------------------------------------------------------------------------
/requirements.txt:
--------------------------------------------------------------------------------
1 | ipython==8.12.3
2 | keras==2.10.0
3 | librosa==0.7.2
4 | matplotlib==3.8.3
5 | numpy==1.24.4
6 | python_speech_features==0.6
7 | scipy==1.13.0
8 | SpeechRecognition==3.10.1
9 | tqdm==4.66.2
10 |
--------------------------------------------------------------------------------
/static/index.css:
--------------------------------------------------------------------------------
1 | * {
2 | box-sizing: border-box;
3 | }
4 |
5 | body {
6 | font-family: "Helvetica Neue", "Roboto", sans-serif;
7 | background-color: #f2f2f2;
8 | margin: 0;
9 | padding: 0;
10 | }
11 |
12 | #header {
13 | background-color: #fff;
14 | color: #333;
15 | display: flex;
16 | justify-content: center;
17 | align-items: center;
18 | height: 80px;
19 | }
20 |
21 | h1 {
22 | font-size: 36px;
23 | margin: 0;
24 | }
25 |
26 | #content {
27 | background-color: #fff;
28 | border-radius: 10px;
29 | box-shadow: 0 2px 10px rgba(0, 0, 0, 0.2);
30 | margin: 50px auto;
31 | max-width: 800px;
32 | padding: 20px;
33 | }
34 |
35 | #content div {
36 | display: flex;
37 | flex-wrap: wrap;
38 | justify-content: space-between;
39 | margin-bottom: 20px;
40 | }
41 |
42 | #content a {
43 | background-color: #fff;
44 | border-radius: 5px;
45 | box-shadow: 0 2px 10px rgba(0, 0, 0, 0.2);
46 | color: #333;
47 | padding: 10px;
48 | text-align: center;
49 | text-decoration: none;
50 | transition: background-color 0.2s;
51 | width: 20%;
52 | }
53 |
54 | #content a:hover {
55 | background-color: #f2f2f2;
56 | }
57 |
58 | #content img {
59 | cursor: pointer;
60 | height: 50px;
61 | transition: transform 0.2s;
62 | width: 50px;
63 | }
64 |
65 | #content img:hover {
66 | transform: scale(1.1);
67 | }
68 |
69 | #result {
70 | background-color: #fff;
71 | border-radius: 5px;
72 | box-shadow: 0 2px 10px rgba(0, 0, 0, 0.2);
73 | padding: 10px;
74 | }
75 |
76 | #result textarea {
77 | border: none;
78 | border-radius: 5px;
79 | font-size: 16px;
80 | height: 300px;
81 | margin-top: 10px;
82 | padding: 10px;
83 | resize: none;
84 | width: 100%;
85 | }
86 |
87 | @media only screen and (max-width: 600px) {
88 | #content a {
89 | width: 100%;
90 | }
91 | }
--------------------------------------------------------------------------------
/static/record.js:
--------------------------------------------------------------------------------
1 | //兼容
2 | window.URL = window.URL || window.webkitURL;
3 | //获取计算机的设备：摄像头或者录音设备
4 | navigator.getUserMedia = navigator.getUserMedia || navigator.webkitGetUserMedia || navigator.mozGetUserMedia || navigator.msGetUserMedia;
5 |
6 | var HZRecorder = function (stream, config) {
7 | config = config || {};
8 | config.sampleBits = config.sampleBits || 16; //采样数位 8, 16
9 | config.sampleRate = config.sampleRate || 16000; //采样率 16000
10 |
11 | //创建一个音频环境对象
12 | var audioContext = window.AudioContext || window.webkitAudioContext;
13 | var context = new audioContext();
14 | var audioInput = context.createMediaStreamSource(stream);
15 | // 第二个和第三个参数指的是输入和输出都是单声道,2是双声道。
16 | var recorder = context.createScriptProcessor(4096, 2, 2);
17 |
18 | var audioData = {
19 | size: 0 //录音文件长度
20 | , buffer: [] //录音缓存
21 | , inputSampleRate: context.sampleRate //输入采样率
22 | , inputSampleBits: 16 //输入采样数位 8, 16
23 | , outputSampleRate: config.sampleRate //输出采样率
24 | , outputSampleBits: config.sampleBits //输出采样数位 8, 16
25 | , input: function (data) {
26 | this.buffer.push(new Float32Array(data));
27 | this.size += data.length;
28 | }
29 | , compress: function () { //合并压缩
30 | //合并
31 | var data = new Float32Array(this.size);
32 | var offset = 0;
33 | for (var i = 0; i < this.buffer.length; i++) {
34 | data.set(this.buffer[i], offset);
35 | offset += this.buffer[i].length;
36 | }
37 | //压缩
38 | var compression = parseInt(this.inputSampleRate / this.outputSampleRate);
39 | var length = data.length / compression;
40 | var result = new Float32Array(length);
41 | var index = 0, j = 0;
42 | while (index < length) {
43 | result[index] = data[j];
44 | j += compression;
45 | index++;
46 | }
47 | return result;
48 | }
49 | , encodeWAV: function () {
50 | var sampleRate = Math.min(this.inputSampleRate, this.outputSampleRate);
51 | var sampleBits = Math.min(this.inputSampleBits, this.outputSampleBits);
52 | var bytes = this.compress();
53 | var dataLength = bytes.length * (sampleBits / 8);
54 | var buffer = new ArrayBuffer(44 + dataLength);
55 | var data = new DataView(buffer);
56 |
57 | var channelCount = 1;//单声道
58 | var offset = 0;
59 |
60 | var writeString = function (str) {
61 | for (var i = 0; i < str.length; i++) {
62 | data.setUint8(offset + i, str.charCodeAt(i));
63 | }
64 | }
65 |
66 | // 资源交换文件标识符
67 | writeString('RIFF');
68 | offset += 4;
69 | // 下个地址开始到文件尾总字节数,即文件大小-8
70 | data.setUint32(offset, 36 + dataLength, true);
71 | offset += 4;
72 | // WAV文件标志
73 | writeString('WAVE');
74 | offset += 4;
75 | // 波形格式标志
76 | writeString('fmt ');
77 | offset += 4;
78 | // 过滤字节,一般为 0x10 = 16
79 | data.setUint32(offset, 16, true);
80 | offset += 4;
81 | // 格式类别 (PCM形式采样数据)
82 | data.setUint16(offset, 1, true);
83 | offset += 2;
84 | // 通道数
85 | data.setUint16(offset, channelCount, true);
86 | offset += 2;
87 | // 采样率,每秒样本数,表示每个通道的播放速度
88 | data.setUint32(offset, sampleRate, true);
89 | offset += 4;
90 | // 波形数据传输率 (每秒平均字节数) 单声道×每秒数据位数×每样本数据位/8
91 | data.setUint32(offset, channelCount * sampleRate * (sampleBits / 8), true);
92 | offset += 4;
93 | // 快数据调整数 采样一次占用字节数 单声道×每样本的数据位数/8
94 | data.setUint16(offset, channelCount * (sampleBits / 8), true);
95 | offset += 2;
96 | // 每样本数据位数
97 | data.setUint16(offset, sampleBits, true);
98 | offset += 2;
99 | // 数据标识符
100 | writeString('data');
101 | offset += 4;
102 | // 采样数据总数,即数据总大小-44
103 | data.setUint32(offset, dataLength, true);
104 | offset += 4;
105 | // 写入采样数据
106 | if (sampleBits === 8) {
107 | for (var i = 0; i < bytes.length; i++, offset++) {
108 | var s = Math.max(-1, Math.min(1, bytes[i]));
109 | var val = s < 0 ? s * 0x8000 : s * 0x7FFF;
110 | val = parseInt(255 / (65535 / (val + 32768)));
111 | data.setInt8(offset, val, true);
112 | }
113 | } else {
114 | for (var i = 0; i < bytes.length; i++, offset += 2) {
115 | var s = Math.max(-1, Math.min(1, bytes[i]));
116 | data.setInt16(offset, s < 0 ? s * 0x8000 : s * 0x7FFF, true);
117 | }
118 | }
119 |
120 | return new Blob([data], {type: 'audio/wav'});
121 | }
122 | };
123 |
124 | //开始录音
125 | this.start = function () {
126 | audioInput.connect(recorder);
127 | recorder.connect(context.destination);
128 | }
129 |
130 | //停止
131 | this.stop = function () {
132 | recorder.disconnect();
133 | }
134 |
135 | //获取音频文件
136 | this.getBlob = function () {
137 | this.stop();
138 | return audioData.encodeWAV();
139 | }
140 |
141 | //回放
142 | this.play = function (audio) {
143 | audio.src = window.URL.createObjectURL(this.getBlob());
144 | }
145 | //清除
146 | this.clear = function () {
147 | audioData.buffer = [];
148 | audioData.size = 0;
149 | }
150 |
151 | //上传
152 | this.upload = function (url, callback) {
153 | var fd = new FormData();
154 | // 上传的文件名和数据
155 | fd.append("audio", this.getBlob());
156 | var xhr = new XMLHttpRequest();
157 | xhr.timeout = 60000
158 | if (callback) {
159 | xhr.upload.addEventListener("progress", function (e) {
160 | callback('uploading', e);
161 | }, false);
162 | xhr.addEventListener("load", function (e) {
163 | callback('ok', e);
164 | }, false);
165 | xhr.addEventListener("error", function (e) {
166 | callback('error', e);
167 | }, false);
168 | xhr.addEventListener("abort", function (e) {
169 | callback('cancel', e);
170 | }, false);
171 | }
172 | xhr.open("POST", url);
173 | xhr.send(fd);
174 | }
175 |
176 | //音频采集
177 | recorder.onaudioprocess = function (e) {
178 | audioData.input(e.inputBuffer.getChannelData(0));
179 | //record(e.inputBuffer.getChannelData(0));
180 | }
181 |
182 | };
183 | //抛出异常
184 | HZRecorder.throwError = function (message) {
185 | alert(message);
186 | throw new function () {
187 | this.toString = function () {
188 | return message;
189 | }
190 | }
191 | }
192 | //是否支持录音
193 | HZRecorder.canRecording = (navigator.getUserMedia != null);
194 | //获取录音机
195 | HZRecorder.get = function (callback, config) {
196 | if (callback) {
197 | if (navigator.getUserMedia) {
198 | navigator.getUserMedia(
199 | {audio: true} //只启用音频
200 | , function (stream) {
201 | var rec = new HZRecorder(stream, config);
202 | callback(rec);
203 | }
204 | , function (error) {
205 | switch (error.code || error.name) {
206 | case 'PERMISSION_DENIED':
207 | case 'PermissionDeniedError':
208 | HZRecorder.throwError('用户拒绝提供信息。');
209 | break;
210 | case 'NOT_SUPPORTED_ERROR':
211 | case 'NotSupportedError':
212 | HZRecorder.throwError('浏览器不支持硬件设备。');
213 | break;
214 | case 'MANDATORY_UNSATISFIED_ERROR':
215 | case 'MandatoryUnsatisfiedError':
216 | HZRecorder.throwError('无法发现指定的硬件设备。');
217 | break;
218 | default:
219 | HZRecorder.throwError('无法打开麦克风。异常信息:' + (error.code || error.name));
220 | break;
221 | }
222 | });
223 | } else {
224 | window.alert('不是HTTPS协议或者localhost地址，不能使用录音功能！')
225 | HZRecorder.throwErr('当前浏览器不支持录音功能。');
226 | return;
227 | }
228 | }
229 | };
--------------------------------------------------------------------------------
/static/record.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/WThirteen/asr_AISHELL-3/3140e6f914dbb3ccac906b40dc6836cdd122e46d/static/record.png
--------------------------------------------------------------------------------
/templates/index.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 | Thirteen_temp
6 |
7 |
8 |
9 |
10 |
13 |
14 |
20 |
21 |
22 |
23 | 上传进度：
24 |
19 |