├── .gitignore ├── LICENSE ├── README.md ├── __init__.py ├── audio ├── __init__.py ├── helpers.py ├── pitch.py └── visuals.py ├── bin ├── __init__.py └── compute_metrics.py ├── config ├── __init__.py └── global_config.py ├── metrics ├── DTW.py ├── FFE.py ├── GPE.py ├── MCD.py ├── MSD.py ├── VDE.py ├── __init__.py ├── dists.py ├── helpers.py └── moments.py └── requirements.txt /.gitignore: -------------------------------------------------------------------------------- 1 | __pycache__ 2 | *.json 3 | .env -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # TTS Objective Metrics 🎯 2 | 3 | This repository comprises a compilation of the objective metrics used in several text-to-speech (TTS) papers. 4 | 5 | ## Available Metrics 6 | | Metric | Used In | 7 | | ------ | ------ | 8 | | Voicing Decision Error (VDE) | [E2E-Prosody](https://arxiv.org/pdf/1803.09047.pdf), [Mellotron](https://arxiv.org/abs/1910.11997)| 9 | | Gross Pitch Error (GPE) | [E2E-Prosody](https://arxiv.org/pdf/1803.09047.pdf), [Mellotron](https://arxiv.org/abs/1910.11997)| 10 | | F0 Frame Error (FFE) | [E2E-Prosody](https://arxiv.org/pdf/1803.09047.pdf), [Mellotron](https://arxiv.org/abs/1910.11997)| 11 | | Dynamic Time Warping (DTW) | [FastSpeech2](https://arxiv.org/abs/2006.04558) | 12 | | Mel Spectral Distortion (MSD) | [Wave-Tacotron](https://arxiv.org/abs/2011.03568) | 13 | | Mel Cepstral Distortion (MCD) | [E2E-Prosody](https://arxiv.org/pdf/1803.09047.pdf), [Wave-Tacotron](https://arxiv.org/abs/2011.03568) | 14 | | Statistical Moments (STD, SKEW, KURT) | [FastSpeech2](https://arxiv.org/abs/2006.04558) | 15 | 16 | ## Available Pitch Computation 17 | | Alogrithm | Proposed In | 18 | | ------ | ------ | 19 | | YIN | [(Cheveigné and Kawahara, 2002)](http://audition.ens.fr/adc/pdf/2002_JASA_YIN.pdf) | 20 | | DIO | [(Morise, Kawahara, and Katayose, 2009)](https://www.aes.org/e-lib/browse.cfm?elib=15165)| 21 | | PYIN (Testing) | [(Mauch and Dixon, 2014)](https://ieeexplore.ieee.org/document/6853678) | 22 | 23 | ## How to Run 24 | First, clone and enter the repo: 25 | ```sh 26 | git clone https://github.com/AI-Unicamp/TTS-Objective-Metrics 27 | cd TTS-Objective-Metrics 28 | ``` 29 | 30 | Install dependencies: 31 | ``` 32 | pip install -r requirements.txt 33 | ``` 34 | 35 | Then, configure the global parameters as you wish in config-> global.py. If not done, all statistics will be computed with the default parameters already set. 36 | 37 | The main usage of the repo is to calculate all available metrics for a batch of (ground truth, synthesized) audio pairs (test/eval set). For this make sure you have the (ground_truth, synthesized) audio pairs names **matching** and in **numbered order, each with the same number of digits of the greatest file** (eg. if there are 100 files, you shall start with 000.wav, if there are [10,99] you shall start with 00.wav). as in: 38 | 39 | 📂My Audio\ 40 | ┣ 📂Ground Truths\ 41 | ┃ ┣ 📜00.wav\ 42 | ┃ ┣ 📜01.wav\ 43 | ┃ ┣ ...\ 44 | ┣ 📂Synthesizeds\ 45 | ┃ ┣ 📜00.wav\ 46 | ┃ ┣ 📜01.wav\ 47 | ┃ ┣ ... 48 | 49 | Then, choose one pitch computing algorithm and run the following command: 50 | ```sh 51 | python -m bin.compute_metrics --gt_folder_path 'My Audio\Ground Truths' --synth_folder_path 'My Audio\Synthesizeds' --pitch_algorithm 'yin' 52 | ``` 53 | The result will be saved in a file named metrics.json in the main repo folder: 'TTS Objective Metrics/metrics.json'. 54 | 55 | Alternatively, it is possible to calculate a single metric for a pair of (ground truth, synthesized) audio, by choosing an available metric and pitch computation method (yin or dio) with one of the following commands: 56 | ```sh 57 | # For DTW, FFE, GPE, VDE, moments 58 | python -m metrics.DTW --gt_path 'ground_truth_audio.wav' --synth_path 'synthesized_audio.wav' --pitch_algorithm 'yin' 59 | ``` 60 | ``` 61 | # For MSD or MCD 62 | python -m metrics.MSD --gt_path 'path_to_ground_truth_audio.wav' --synth_path 'path_to_synthesized_audio.wav' 63 | ``` 64 | The result will be displayed in the terminal. 65 | 66 | ## Repo Organization 67 | 📦TTS Objective Metrics\ 68 | ┣ 📂audio\ 69 | ┃ ┣ 📜helpers.py\ 70 | ┃ ┣ 📜pitch.py\ 71 | ┃ ┣ 📜visuals.py\ 72 | ┣ 📂bin\ 73 | ┃ ┣ 📜compute_metrics.py\ 74 | ┣ 📂config\ 75 | ┃ ┣ 📜global_config.py\ 76 | ┣ 📂metrics\ 77 | ┃ ┣ 📜dists.py\ 78 | ┃ ┣ 📜DTW.py\ 79 | ┃ ┣ 📜FFE.py\ 80 | ┃ ┣ 📜GPE.py\ 81 | ┃ ┣ 📜helpers.py\ 82 | ┃ ┣ 📜MCD.py\ 83 | ┃ ┣ 📜moments.py\ 84 | ┃ ┣ 📜MSD.py\ 85 | ┃ ┣ 📜VDE.py\ 86 | ┣ 📜README.md 87 | 88 | ## How to Contribute 89 | As the repo is still in its infancy, feel free to either open an issue, discussion or send a pull request, or even contact us by e-mail. 90 | 91 | ## Authors 92 | - Leonardo B. de M. M. Marques (leonardoboulitreau@gmail.com) 93 | - Lucas Hideki Ueda (lucashueda@gmail.com) 94 | 95 | ## Github references 96 | - [Coqui-AI](https://github.com/coqui-ai/TTS) 97 | - [Facebook Fairseq](https://github.com/pytorch/fairseq) 98 | - [NVIDIA Deep Learning Examples](https://github.com/NVIDIA/DeepLearningExamples) 99 | - [NVIDIA Mellotron](https://github.com/NVIDIA/mellotron/tree/d5362ccae23984f323e3cb024a01ec1de0493aff) 100 | - [MAPS](https://github.com/bastibe/MAPS-Scripts) 101 | - [Yin](https://github.com/patriceguyot/Yin) 102 | 103 | All references are listened on top of the used code itself. 104 | -------------------------------------------------------------------------------- /__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AI-Unicamp/TTS-Objective-Metrics/e183fd0f4ec484173b13b86ae315c862af9db888/__init__.py -------------------------------------------------------------------------------- /audio/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AI-Unicamp/TTS-Objective-Metrics/e183fd0f4ec484173b13b86ae315c862af9db888/audio/__init__.py -------------------------------------------------------------------------------- /audio/helpers.py: -------------------------------------------------------------------------------- 1 | import os 2 | import glob 3 | import librosa 4 | 5 | # Load Audios 6 | def read_folder(path): 7 | out = list() 8 | for filename in sorted(glob.glob(os.path.join(path, '*.wav'))): 9 | x, sr = librosa.load(filename) 10 | out.append(x) 11 | return out, sr -------------------------------------------------------------------------------- /audio/pitch.py: -------------------------------------------------------------------------------- 1 | import librosa 2 | import numpy as np 3 | import torch 4 | import pyworld as pw 5 | 6 | # https://github.com/pytorch/fairseq/blob/fcca32258c8e8bcc9f9890bf4714fa2f96b6b3e1/examples/speech_synthesis/evaluation/eval_f0.py#L87 7 | def yin(sig, config): 8 | 9 | '''Compute the Yin Algorithm. Return fundamental frequency and harmonic rate. 10 | An adaptation of https://github.com/NVIDIA/mellotron which is an adaption of 11 | https://github.com/patriceguyot/Yin 12 | 13 | Args: 14 | sig (list of floats): Audio Signal 15 | sr (int): Sample Rate 16 | win_len (int): Size of the analysis window (samples) 17 | hop_length (int): Size of the lag between two consecutives windows (samples) 18 | f0_min (int): Minimum fundamental frequency that can be detected (hertz) 19 | f0_max (int): Maximum fundamental frequency that can be detected (hertz) 20 | harmo_thresh (float): Threshold of detection. The algorithm returns the 21 | first minimum of the CMND function below this threshold. 22 | 23 | Returns: 24 | (dict) with: 25 | 'pitches' (1-D np.array): fundamental frequencies, 26 | 'harmonic_rates' (1-D np.array): list of harmonic rate values for each 27 | fundamental frequency value (= confidence value) 28 | 'argmins' (1-D np.array): minimums of the Cumulative Mean Normalized 29 | DifferenceFunction 30 | 'times' (1-D np.array): list of time of each estimation 31 | ''' 32 | sr = config.sr 33 | win_length = config.win_length 34 | hop_length = config.hop_length 35 | f0_min = config.f0_min_pitch 36 | f0_max = config.f0_max_pitch 37 | harmo_thresh = config.harmo_thresh 38 | 39 | tau_min = int(sr / f0_max) 40 | tau_max = int(sr / f0_min) 41 | 42 | # Time values for each analysis window 43 | time_scale = range(0, len(sig) - win_length, hop_length) 44 | times = [t/float(sr) for t in time_scale] 45 | frames = [sig[t:t + win_length] for t in time_scale] 46 | 47 | pitches = [0.0] * len(time_scale) 48 | harmonic_rates = [0.0] * len(time_scale) 49 | argmins = [0.0] * len(time_scale) 50 | 51 | for i, frame in enumerate(frames): 52 | # Compute YIN 53 | df = _difference_function(frame, win_length, tau_max) 54 | cm_df = _cumulative_mean_normalized_difference_function(df, tau_max) 55 | p = _get_pitch(cm_df, tau_min, tau_max, harmo_thresh) 56 | 57 | # Get results 58 | if np.argmin(cm_df) > tau_min: 59 | argmins[i] = float(sr / np.argmin(cm_df)) 60 | if p != 0: # A pitch was found 61 | pitches[i] = float(sr / p) 62 | harmonic_rates[i] = cm_df[p] 63 | else: # No pitch, but we compute a value of the harmonic rate 64 | harmonic_rates[i] = min(cm_df) 65 | 66 | return {'pitches' : np.array(pitches), 'harmonic_rates' : np.array(harmonic_rates), 'argmins' : np.array(argmins), 'times' : np.array(times)} 67 | 68 | def _difference_function(x, n, tau_max): 69 | 70 | '''Compute difference function of data x. This solution is implemented directly 71 | with Numpy fft. 72 | 73 | Args: 74 | x (list): Audio data frame 75 | n (int): Length of data 76 | tau_max (int): Integration window size 77 | 78 | Returns: 79 | (list) of difference function 80 | ''' 81 | x = np.array(x, np.float64) 82 | w = x.size 83 | tau_max = min(tau_max, w) 84 | x_cumsum = np.concatenate((np.array([0.]), (x * x).cumsum())) 85 | size = w + tau_max 86 | p2 = (size // 32).bit_length() 87 | nice_numbers = (16, 18, 20, 24, 25, 27, 30, 32) 88 | size_pad = min(x * 2 ** p2 for x in nice_numbers if x * 2 ** p2 >= size) 89 | fc = np.fft.rfft(x, size_pad) 90 | conv = np.fft.irfft(fc * fc.conjugate())[:tau_max] 91 | return x_cumsum[w:w - tau_max:-1] + x_cumsum[w] - x_cumsum[:tau_max] - \ 92 | 2 * conv 93 | 94 | def _cumulative_mean_normalized_difference_function(df, n): 95 | 96 | '''Compute cumulative mean normalized difference function (CMND). 97 | 98 | Args: 99 | df (list): Difference function 100 | n (int): Length of data 101 | 102 | Returns: 103 | (list) of cumulative mean normalized difference function 104 | ''' 105 | 106 | # Scipy method 107 | cmn_df = df[1:] * range(1, n) / np.cumsum(df[1:]).astype(float) 108 | return np.insert(cmn_df, 0, 1) 109 | 110 | def _get_pitch(cmdf, tau_min, tau_max, harmo_th=0.1): 111 | 112 | '''Return fundamental period of a frame based on CMND function. 113 | 114 | Args: 115 | cmdf (list): Cumulative Mean Normalized Difference function 116 | tau_min (int): Minimum period for speech 117 | tau_max (int): Maximum period for speech 118 | harmo_th (float): harmonicity threshold to determine if 119 | it is necessary to compute pitch frequency 120 | 121 | Returns: 122 | (float) fundamental period if there is values under threshold, 123 | 0 otherwise 124 | ''' 125 | 126 | tau = tau_min 127 | while tau < tau_max: 128 | if cmdf[tau] < harmo_th: 129 | while tau + 1 < tau_max and cmdf[tau + 1] < cmdf[tau]: 130 | tau += 1 131 | return tau 132 | tau += 1 133 | 134 | return 0 # if unvoiced 135 | 136 | # https://github.com/coqui-ai/TTS/blob/main/TTS/utils/audio.py#L710 137 | def dio(sig, config): 138 | '''Compute pitch (f0) of a waveform using the same parameters used for 139 | computing melspectrogram. 140 | 141 | Args: 142 | sig (list of floats): Audio Signal 143 | sr (int): Sample Rate 144 | hop_length (int): Number of frames between STFT columns 145 | f0_min (int): Minimum fundamental frequency that can be detected (hertz) 146 | f0_max (int): Maximum fundamental frequency that can be detected (hertz) 147 | 148 | Returns: 149 | (dict) with: 150 | 'pitches' (1-D np.array): fundamental frequencies, 151 | 'times' (1-D np.array): list of time of each estimation 152 | ''' 153 | sr = config.sr 154 | hop_length = config.hop_length 155 | f0_min = config.f0_min_pitch 156 | f0_max = config.f0_max_pitch 157 | 158 | 159 | # Align F0 length to the spectrogram length 160 | if len(sig) % hop_length == 0: 161 | sig = np.pad(sig, (0, hop_length // 2), mode="reflect") 162 | 163 | f0, t = pw.dio( 164 | sig.astype(np.double), 165 | fs=sr, 166 | f0_ceil=f0_max, 167 | f0_floor = f0_min, 168 | frame_period=1000 * hop_length / sr, 169 | ) 170 | f0 = pw.stonemask(sig.astype(np.double), f0, t, sr) 171 | return {'pitches':f0, 'times':t} 172 | 173 | # NOT TESTED YET 174 | # https://github.com/NVIDIA/DeepLearningExamples/blob/a43ffd01cb002b23a98c97c3c5a231e24a57fa71/PyTorch/SpeechSynthesis/FastPitch/fastpitch/data_function.py#L81 175 | def compute_pyin(wav, mel_len, method='pyin', normalize_mean=None,normalize_std=None, n_formants=1): 176 | 177 | if type(normalize_mean) is float or type(normalize_mean) is list: 178 | normalize_mean = torch.tensor(normalize_mean) 179 | 180 | if type(normalize_std) is float or type(normalize_std) is list: 181 | normalize_std = torch.tensor(normalize_std) 182 | 183 | if method == 'pyin': 184 | 185 | snd, sr = librosa.load(wav) 186 | pitch_mel, voiced_flag, voiced_probs = librosa.pyin( 187 | snd, fmin=librosa.note_to_hz('C2'), 188 | fmax=librosa.note_to_hz('C7'), frame_length=1024) 189 | assert np.abs(mel_len - pitch_mel.shape[0]) <= 1.0 190 | 191 | pitch_mel = np.where(np.isnan(pitch_mel), 0.0, pitch_mel) 192 | pitch_mel = torch.from_numpy(pitch_mel).unsqueeze(0) 193 | pitch_mel = torch.nn.functional.pad(pitch_mel, (0, mel_len - pitch_mel.size(1))) 194 | 195 | if n_formants > 1: 196 | raise NotImplementedError 197 | 198 | else: 199 | raise ValueError 200 | 201 | pitch_mel = pitch_mel.float() 202 | 203 | if normalize_mean is not None: 204 | assert normalize_std is not None 205 | pitch_mel = _normalize_pitch(pitch_mel, normalize_mean, normalize_std) 206 | 207 | return pitch_mel 208 | 209 | def _normalize_pitch(pitch, mean, std): 210 | zeros = (pitch == 0.0) 211 | pitch -= mean[:, None] 212 | pitch /= std[:, None] 213 | pitch[zeros] = 0.0 214 | return pitch -------------------------------------------------------------------------------- /audio/visuals.py: -------------------------------------------------------------------------------- 1 | import matplotlib.pyplot as plt 2 | import numpy as np 3 | import torch 4 | 5 | # https://github.com/coqui-ai/TTS/blob/0592a5805ca7fb877c8fc8df56a4eacac13f2657/TTS/tts/utils/visual.py#L35 6 | def plot_spectrogram(spectrogram, fig_size): 7 | '''Plot spectrogram. 8 | 9 | Args: 10 | spectrogram (np.array): Spectrogram values. 11 | 12 | Returns: 13 | plt.figure 14 | ''' 15 | if isinstance(spectrogram, torch.Tensor): 16 | spectrogram_ = spectrogram.detach().cpu().numpy().squeeze().T 17 | else: 18 | spectrogram_ = spectrogram.T 19 | spectrogram_ = spectrogram_.astype(np.float32) if spectrogram_.dtype == np.float16 else spectrogram_ 20 | fig = plt.figure(figsize=fig_size) 21 | plt.imshow(spectrogram_, aspect="auto", origin="lower") 22 | plt.colorbar() 23 | plt.tight_layout() 24 | plt.close() 25 | return fig 26 | 27 | # https://github.com/coqui-ai/TTS/blob/0592a5805ca7fb877c8fc8df56a4eacac13f2657/TTS/tts/utils/visual.py#L52 28 | def plot_pitch(pitch, spectrogram, fig_size): 29 | '''Plot pitch curves on top of the spectrogram. 30 | 31 | Args: 32 | pitch ( (T,)np.array): Pitch values. 33 | spectrogram ( (C, T) np.array): Spectrogram values. 34 | 35 | Returns: 36 | plt.figure 37 | ''' 38 | 39 | if isinstance(spectrogram, torch.Tensor): 40 | spectrogram_ = spectrogram.detach().cpu().numpy().squeeze().T 41 | else: 42 | spectrogram_ = spectrogram.T 43 | spectrogram_ = spectrogram_.astype(np.float32) if spectrogram_.dtype == np.float16 else spectrogram_ 44 | 45 | old_fig_size = plt.rcParams["figure.figsize"] 46 | if fig_size is not None: 47 | plt.rcParams["figure.figsize"] = fig_size 48 | 49 | fig, ax = plt.subplots() 50 | 51 | ax.imshow(spectrogram_, aspect="auto", origin="lower") 52 | ax.set_xlabel("time") 53 | ax.set_ylabel("spec_freq") 54 | 55 | ax2 = ax.twinx() 56 | ax2.plot(pitch, linewidth=5.0, color="red") 57 | ax2.set_ylabel("F0") 58 | 59 | plt.rcParams["figure.figsize"] = old_fig_size 60 | plt.close() 61 | return fig 62 | 63 | 64 | # Testing Plotting DTW Alignment between pitchs/spectrograms 65 | #def plot_aligned_pitch(x, y, path, sample_init=0, sample_end=400): 66 | # 67 | # x = x[sample_init:sample_end] 68 | # y = y[sample_init:sample_end] 69 | # path = path[sample_init:sample_end] 70 | 71 | # plt.figure(figsize=(10, 8)) 72 | # dis_y = 150 73 | # dis_x = 0 74 | # plt.plot(np.arange(x.shape[0]) + dis_x, x + dis_y, "-", c="C3", linewidth = 3) 75 | # plt.plot(np.arange(y.shape[0]) - dis_x, y - dis_y, "-", c="C0", linewidth = 3) 76 | # for x_i, y_j in path: 77 | # plt.plot([x_i + dis_x, y_j - dis_x], [x[x_i] + dis_y, y[y_j] - dis_y], "-", c="C7", linewidth = 0.5) 78 | # plt.axis("off") 79 | # plt.savefig("mygraph.png") 80 | 81 | # a = [] 82 | # b = [] 83 | # a.append(torch.Tensor(refs_yin[0][0]).unsqueeze_(1)) 84 | # b.append(torch.Tensor(synths_yin[0][0]).unsqueeze_(1)) 85 | # x,y,z,k = batch_dynamic_time_warping(a,b, _compute_rms_dist) 86 | # plot_aligned_pitch(np.array(refs_yin[0][0]), np.array(synths_yin[0][0]), k) -------------------------------------------------------------------------------- /bin/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AI-Unicamp/TTS-Objective-Metrics/e183fd0f4ec484173b13b86ae315c862af9db888/bin/__init__.py -------------------------------------------------------------------------------- /bin/compute_metrics.py: -------------------------------------------------------------------------------- 1 | # Basic Imports 2 | import argparse 3 | import json 4 | import numpy as np 5 | import torch 6 | 7 | # Import Audio Loader 8 | from audio.helpers import read_folder 9 | 10 | # Import Pitch Computation 11 | from audio.pitch import dio, yin 12 | 13 | # Import Config 14 | from config.global_config import GlobalConfig 15 | 16 | # Import Metrics 17 | from metrics.VDE import voicing_decision_error 18 | from metrics.GPE import gross_pitch_error 19 | from metrics.FFE import f0_frame_error 20 | from metrics.DTW import batch_dynamic_time_warping 21 | from metrics.MSD import batch_mel_spectral_distortion 22 | from metrics.MCD import batch_mel_cepstral_distortion 23 | from metrics.moments import estimate_moments 24 | 25 | # Import Basic Stats 26 | from metrics.helpers import add_basic_stats 27 | 28 | # Get Input Args 29 | parser = argparse.ArgumentParser(description='Script to calculate the all repo metrics between the ground truth and the synthesized audios.') 30 | parser.add_argument("--gt_folder_path", required = True, type=str, help = 'Path to the folder containing all ground truth audio in .wav format') 31 | parser.add_argument("--synth_folder_path", required = True, type=str, help = 'Path to the folder containing all respective synthesized audio in .wav format') 32 | parser.add_argument("--pitch_algorithm", required = True, type=str, choices = ["dio", "yin"], help = 'Choose method of computing pitch') 33 | args = parser.parse_args() 34 | gt_folder_path = args.gt_folder_path 35 | synth_folder_path = args.synth_folder_path 36 | pitch_algorithm = args.pitch_algorithm 37 | 38 | # Import Audios 39 | refs, _ = read_folder(gt_folder_path) 40 | synths, _ = read_folder(synth_folder_path) 41 | refs_tensor = [torch.Tensor(item) for item in refs] 42 | synths_tensor = [torch.Tensor(item) for item in synths] 43 | 44 | # Define Configurations 45 | config = GlobalConfig() 46 | 47 | # Compute Pitch 48 | refs_pitch = [eval(pitch_algorithm)(item, config) for item in refs] 49 | synths_pitch = [eval(pitch_algorithm)(item, config) for item in synths] 50 | refs_pitch_tensor = [torch.Tensor(item['pitches']).unsqueeze_(1) for item in refs_pitch] 51 | synths_pitch_tensor = [torch.Tensor(item['pitches']).unsqueeze_(1) for item in synths_pitch] 52 | 53 | # Pitch DistributionS 54 | refs_dist = [item for sublist in refs_pitch for item in sublist['pitches']] 55 | synths_dist = [item for sublist in synths_pitch for item in sublist['pitches']] 56 | 57 | # Pairwise Metrics 58 | VDE = {} 59 | GPE = {} 60 | FFE = {} 61 | for i in range(len(refs)): 62 | VDE[i] = voicing_decision_error(refs_pitch[i]['times'], refs_pitch[i]['pitches'], synths_pitch[i]['times'], synths_pitch[i]['pitches']) 63 | GPE[i] = gross_pitch_error(refs_pitch[i]['times'], refs_pitch[i]['pitches'], synths_pitch[i]['times'], synths_pitch[i]['pitches']) 64 | FFE[i] = f0_frame_error(refs_pitch[i]['times'], refs_pitch[i]['pitches'], synths_pitch[i]['times'], synths_pitch[i]['pitches']) 65 | 66 | # Batched Metrics 67 | DTW = {v:k for v,k in enumerate(batch_dynamic_time_warping(refs_pitch_tensor, synths_pitch_tensor, config.dist_fn, config.norm_align_type)['norm_align_costs'])} 68 | MSD = {v:k for v,k in enumerate(batch_mel_spectral_distortion(refs_tensor, synths_tensor, config))} 69 | MCD = {v:k for v,k in enumerate(batch_mel_cepstral_distortion(refs_tensor, synths_tensor, config))} 70 | 71 | # Distribution Metrics 72 | MOMENTS = estimate_moments(refs_dist, synths_dist) 73 | 74 | # Compute Basic Stats and Store 75 | stats = {} 76 | stats['VDE'] = add_basic_stats(VDE) 77 | stats['GPE'] = add_basic_stats(GPE) 78 | stats['FFE'] = add_basic_stats(FFE) 79 | stats['DTW'] = add_basic_stats(DTW) 80 | stats['MSD'] = add_basic_stats(MSD) 81 | stats['MCD'] = add_basic_stats(MCD) 82 | stats['Moments'] = MOMENTS 83 | 84 | # Create Output File 85 | tts_objective_metrics = {'tts_objecive_metrics': stats} 86 | json_object = json.dumps(tts_objective_metrics, indent = 4) 87 | with open("metrics.json", "w") as outfile: 88 | outfile.write(json_object) 89 | -------------------------------------------------------------------------------- /config/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AI-Unicamp/TTS-Objective-Metrics/e183fd0f4ec484173b13b86ae315c862af9db888/config/__init__.py -------------------------------------------------------------------------------- /config/global_config.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torchaudio 3 | 4 | class GlobalConfig(): 5 | def __init__(self): 6 | 7 | self.sr = 22050 8 | 9 | self.f0_min_pitch: int = 100 10 | self.f0_max_pitch: int = 500 11 | 12 | self.n_fft = 1024 13 | self.win_length: int = 1024 14 | self.hop_length: int = 256 15 | self.f0_min_mel: int = 0 16 | self.f0_max_mel = None 17 | self.harmo_thresh: float = 0.1 18 | self.window_fn = torch.hann_window 19 | self.log_mels = True 20 | self.n_mels = 80 21 | self.n_mfcc = 13 22 | self.power = 2 23 | self.fig_size: tuple = (16,10) 24 | self.dist_fn = 'compute_rms_dist' 25 | self.norm_align_type = 'path' 26 | 27 | self.mel_fn = torchaudio.transforms.MelSpectrogram( 28 | sample_rate = self.sr, n_fft=self.n_fft, win_length=self.win_length, 29 | hop_length=self.hop_length, f_min=self.f0_min_mel, f_max = self.f0_max_mel, 30 | n_mels=self.n_mels, window_fn=self.window_fn, power = self.power 31 | ) 32 | 33 | self.melkwargs = { 34 | "n_fft": self.n_fft, 35 | "win_length": self.win_length, 36 | "hop_length": self.hop_length, 37 | "f_min": self.f0_min_mel, 38 | "f_max": self.f0_max_mel, 39 | "window_fn": self.window_fn, 40 | "n_mels": self.n_mels, 41 | "power": self.power 42 | } 43 | 44 | self.mfcc_fn = torchaudio.transforms.MFCC( 45 | sample_rate = self.sr, n_mfcc=self.n_mfcc, log_mels=self.log_mels, melkwargs=self.melkwargs 46 | ) 47 | -------------------------------------------------------------------------------- /metrics/DTW.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import librosa 3 | import numpy as np 4 | import torch 5 | from config.global_config import GlobalConfig 6 | from audio.pitch import dio, yin 7 | from metrics.dists import compute_rms_dist 8 | 9 | # https://github.com/pytorch/fairseq/blob/fcca32258c8e8bcc9f9890bf4714fa2f96b6b3e1/fairseq/tasks/text_to_speech.py#L304 10 | def batch_dynamic_time_warping(x1, x2, dist_fn, norm_align_type): 11 | """full batched DTW without any constraints 12 | x1: list of tensors of the references 13 | x2: list of tensors of the synthesizeds 14 | dist_fn: distance metric used to compute 15 | outputs: list of floats/ints or list of numpy arrays indexed by batch 16 | """ 17 | 18 | distance, shapes = [], [] 19 | for cur_x1, cur_x2 in zip(x1, x2): 20 | # Compute Distance between the Ref/Synths 21 | cur_d = eval(dist_fn)(cur_x1, cur_x2) 22 | # cur_d = dist_fn(cur_x1, cur_x2) 23 | distance.append(cur_d) 24 | shapes.append(distance[-1].size()) 25 | 26 | # Get Max Spec Size of References 27 | max_m = max(ss[0] for ss in shapes) 28 | # Get Max Spec Size of Synths 29 | max_n = max(ss[1] for ss in shapes) 30 | 31 | # Pad all distance matrices with to be [Max_Synth, Max_Ref] 32 | distance = torch.stack( 33 | [torch.nn.functional.pad(dd, (0, max_n - dd.size(1), 0, max_m - dd.size(0))) for dd in distance] 34 | ) 35 | 36 | # All Spec Sizes (List of [REF_SPEC_SIZE, SYNTH_SPEC_SIZE]) 37 | shapes = torch.LongTensor(shapes).to(distance.device) 38 | 39 | # ptr: 0=left, 1=up-left, 2=up 40 | ptr2dij = {0: (0, -1), 1: (-1, -1), 2: (-1, 0)} 41 | 42 | bsz, m, n = distance.size() 43 | cumdist = torch.zeros_like(distance) 44 | backptr = torch.zeros_like(distance).type(torch.int32) - 1 45 | 46 | # initialize 47 | cumdist[:, 0, :] = distance[:, 0, :].cumsum(dim=-1) 48 | cumdist[:, :, 0] = distance[:, :, 0].cumsum(dim=-1) 49 | backptr[:, 0, :] = 0 50 | backptr[:, :, 0] = 2 51 | 52 | # DP with optimized anti-diagonal parallelization, O(M+N) steps 53 | for offset in range(2, m + n - 1): 54 | ind = _antidiag_indices(offset, 1, m, 1, n) 55 | c = torch.stack( 56 | [ 57 | cumdist[:, ind[0], ind[1] - 1], 58 | cumdist[:, ind[0] - 1, ind[1] - 1], 59 | cumdist[:, ind[0] - 1, ind[1]], 60 | ], 61 | dim=2, 62 | ) 63 | v, b = c.min(axis=-1) 64 | 65 | # Best Backpaths from each starting point 66 | backptr[:, ind[0], ind[1]] = b.int() 67 | 68 | # D(i,j) = d(x_i,y_i) + min(D(i-1,j-1), D(i-1,j), D(i,j-1)) 69 | cumdist[:, ind[0], ind[1]] = v + distance[:, ind[0], ind[1]] 70 | 71 | # Evaluate the Optimal Backpath 72 | pathmap = torch.zeros_like(backptr) 73 | for b in range(bsz): 74 | i = m - 1 if shapes is None else (shapes[b][0] - 1).item() 75 | j = n - 1 if shapes is None else (shapes[b][1] - 1).item() 76 | dtwpath = [(i, j)] 77 | while (i != 0 or j != 0) and len(dtwpath) < 10000: 78 | assert i >= 0 and j >= 0 79 | di, dj = ptr2dij[backptr[b, i, j].item()] 80 | i, j = i + di, j + dj 81 | dtwpath.append((i, j)) 82 | dtwpath = dtwpath[::-1] 83 | indices = torch.from_numpy(np.array(dtwpath)).long() 84 | pathmap[b, indices[:, 0], indices[:, 1]] = 1 85 | 86 | norm_align_costs = [] 87 | cumdists = [] 88 | backptrs = [] 89 | pathmaps = [] 90 | idxs = [] 91 | itr = zip(shapes, cumdist, backptr, pathmap) 92 | 93 | # Get distortion for each item in the batch 94 | for (m, n), cumdist, backptr, pathmap in itr: 95 | cumdist = cumdist[:m, :n] 96 | backptr = backptr[:m, :n] 97 | pathmap = pathmap[:m, :n] 98 | divisor = _get_divisor(pathmap, norm_align_type) 99 | norm_align_cost = cumdist[-1, -1] / divisor 100 | 101 | # Get Indices 102 | p = pathmap.squeeze(0).cpu().detach().numpy() 103 | idx = np.transpose(np.nonzero(p)) 104 | 105 | norm_align_costs.append(float(norm_align_cost.item())) 106 | cumdists.append(cumdist.numpy()) 107 | backptrs.append(backptr.numpy()) 108 | pathmaps.append(pathmap.numpy()) 109 | idxs.append(idx) 110 | 111 | return {'norm_align_costs' : norm_align_costs, 'cumdists' : cumdists, 'backptrs' : backptrs, 'pathmaps' : pathmaps, 'idxs': idxs} 112 | 113 | def _antidiag_indices(offset, min_i=0, max_i=None, min_j=0, max_j=None): 114 | """ 115 | for a (3, 4) matrix with min_i=1, max_i=3, min_j=1, max_j=4, outputs 116 | offset=2 (1, 1), 117 | offset=3 (2, 1), (1, 2) 118 | offset=4 (2, 2), (1, 3) 119 | offset=5 (2, 3) 120 | constraints: 121 | i + j = offset 122 | min_j <= j < max_j 123 | min_i <= offset - j < max_i 124 | """ 125 | if max_i is None: 126 | max_i = offset + 1 127 | if max_j is None: 128 | max_j = offset + 1 129 | min_j = max(min_j, offset - max_i + 1, 0) 130 | max_j = min(max_j, offset - min_i + 1, offset + 1) 131 | j = torch.arange(min_j, max_j) 132 | i = offset - j 133 | return torch.stack([i, j]) 134 | 135 | def _get_divisor(pathmap, normalize_type): 136 | if normalize_type is None: 137 | return 1 138 | elif normalize_type == "len1": 139 | return pathmap.size(0) 140 | elif normalize_type == "len2": 141 | return pathmap.size(1) 142 | elif normalize_type == "path": 143 | return pathmap.sum().item() 144 | else: 145 | raise ValueError(f"normalize_type {normalize_type} not supported") 146 | 147 | def main(gt_path, synth_path, pitch_algorithm): 148 | 149 | # Load Audio 150 | x_gt, _ = librosa.load(gt_path) 151 | x_synth, _ = librosa.load(synth_path) 152 | 153 | # Load Config 154 | config = GlobalConfig() 155 | 156 | # Compute Pitch 157 | pitch_gt = eval(pitch_algorithm)(x_gt, config) 158 | pitch_synth = eval(pitch_algorithm)(x_synth, config) 159 | 160 | # Initiate Lists of Tensors for Batched DTW 161 | x1 = [] 162 | x2 = [] 163 | 164 | # Append Each Tensor 165 | x1.append(torch.Tensor(pitch_gt['pitches']).unsqueeze_(1)) 166 | x2.append(torch.Tensor(pitch_synth['pitches']).unsqueeze_(1)) 167 | 168 | return print(batch_dynamic_time_warping(x1, x2, config.dist_fn, config.norm_align_type)['norm_align_costs']) 169 | 170 | if __name__ == "__main__": 171 | 172 | parser = argparse.ArgumentParser(description='Script to calculate the Dynamic Time Warping between the ground truth and the synthesized audios.') 173 | parser.add_argument("--gt_path", required = True, type=str, help = 'Path to corresponding ground truth audio in .wav format') 174 | parser.add_argument("--synth_path", required = True, type=str, help = 'Path to corresponding synthesized audio in .wav format') 175 | parser.add_argument("--pitch_algorithm", required = True, type=str, choices = ["dio", "yin"], help = 'Choose method of computing pitch') 176 | args = parser.parse_args() 177 | 178 | gt_path = args.gt_path 179 | synth_path = args.synth_path 180 | pitch_algorithm = args.pitch_algorithm 181 | 182 | main(gt_path, synth_path, pitch_algorithm) -------------------------------------------------------------------------------- /metrics/FFE.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import librosa 3 | import numpy as np 4 | from audio.pitch import yin, dio 5 | from config.global_config import GlobalConfig 6 | from metrics.helpers import same_t_in_true_and_est, gross_pitch_error_frames, voicing_decision_error_frames 7 | 8 | # https://github.com/bastibe/MAPS-Scripts/blob/master/helper.py 9 | @same_t_in_true_and_est 10 | def f0_frame_error(true_t, true_f, est_t, est_f): 11 | """Measurs the percentage of frames that either are outside a threshold 12 | around the true pitch or that contain a voicing decision error. 13 | """ 14 | gpe_frames = gross_pitch_error_frames( 15 | true_t, true_f, est_t, est_f 16 | ) 17 | vde_frames = voicing_decision_error_frames( 18 | true_t, true_f, est_t, est_f 19 | ) 20 | return (np.sum(gpe_frames) + 21 | np.sum(vde_frames)) / (len(true_t)) 22 | 23 | def main(gt_path, synth_path, pitch_algorithm): 24 | 25 | # Load Audio 26 | x_gt, _ = librosa.load(gt_path) 27 | x_synth, _ = librosa.load(synth_path) 28 | 29 | # Load Config 30 | config = GlobalConfig() 31 | 32 | # Compute Pitch 33 | pitch_gt = eval(pitch_algorithm)(x_gt, config) 34 | pitch_synth = eval(pitch_algorithm)(x_synth, config) 35 | 36 | return print(f0_frame_error(np.array(pitch_gt['times']), np.array(pitch_gt['pitches']), np.array(pitch_synth['times']), np.array(pitch_synth['pitches']))) 37 | 38 | if __name__ == "__main__": 39 | 40 | parser = argparse.ArgumentParser(description='Script to calculate the F0 Frame Error between the ground truth and the synthesized audios.') 41 | parser.add_argument("--gt_path", required = True, type=str, help = 'Path to corresponding ground truth audio in .wav format') 42 | parser.add_argument("--synth_path", required = True, type=str, help = 'Path to corresponding synthesized audio in .wav format') 43 | parser.add_argument("--pitch_algorithm", required = True, type=str, choices = ["dio", "yin", "pyin"], help = 'Choose method of computing pitch') 44 | 45 | args = parser.parse_args() 46 | 47 | gt_path = args.gt_path 48 | synth_path = args.synth_path 49 | pitch_algorithm = args.pitch_algorithm 50 | 51 | main(gt_path, synth_path, pitch_algorithm) -------------------------------------------------------------------------------- /metrics/GPE.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import librosa 3 | import numpy as np 4 | from audio.pitch import yin, dio 5 | from config.global_config import GlobalConfig 6 | from metrics.helpers import true_voiced_frames, gross_pitch_error_frames, same_t_in_true_and_est 7 | 8 | # https://github.com/bastibe/MAPS-Scripts/blob/master/helper.py 9 | @same_t_in_true_and_est 10 | def gross_pitch_error(true_t, true_f, est_t, est_f): 11 | """The relative frequency in percent of pitch estimates that are 12 | outside a threshold around the true pitch. Only frames that are 13 | considered pitched by both the ground truth and the estimator (if 14 | applicable) are considered. 15 | """ 16 | correct_frames = true_voiced_frames(true_t, true_f, est_t, est_f) 17 | gpe_frames = gross_pitch_error_frames(true_t, true_f, est_t, est_f) 18 | return np.sum(gpe_frames) / np.sum(correct_frames) 19 | 20 | def main(gt_path, synth_path, pitch_algorithm): 21 | 22 | # Load Audio 23 | x_gt, _ = librosa.load(gt_path) 24 | x_synth, _ = librosa.load(synth_path) 25 | 26 | # Load Config 27 | config = GlobalConfig() 28 | 29 | # Compute Pitch 30 | pitch_gt = eval(pitch_algorithm)(x_gt, config) 31 | pitch_synth = eval(pitch_algorithm)(x_synth, config) 32 | 33 | return print(gross_pitch_error(np.array(pitch_gt['times']), np.array(pitch_gt['pitches']), np.array(pitch_synth['times']), np.array(pitch_synth['pitches']))) 34 | 35 | if __name__ == "__main__": 36 | 37 | parser = argparse.ArgumentParser(description='Script to calculate the Gross Pitch Error between the ground truth and the synthesized audios.') 38 | parser.add_argument("--gt_path", required = True, type=str, help = 'Path to corresponding ground truth audio in .wav format') 39 | parser.add_argument("--synth_path", required = True, type=str, help = 'Path to corresponding synthesized audio in .wav format') 40 | parser.add_argument("--pitch_algorithm", required = True, type=str, choices = ["dio", "yin"], help = 'Choose method of computing pitch') 41 | 42 | args = parser.parse_args() 43 | 44 | gt_path = args.gt_path 45 | synth_path = args.synth_path 46 | pitch_algorithm = args.pitch_algorithm 47 | 48 | main(gt_path, synth_path, pitch_algorithm) -------------------------------------------------------------------------------- /metrics/MCD.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | from config.global_config import GlobalConfig 3 | import librosa 4 | import torch 5 | import torchaudio 6 | from metrics.helpers import batch_compute_distortion 7 | 8 | # https://github.com/pytorch/fairseq/blob/fcca32258c8e8bcc9f9890bf4714fa2f96b6b3e1/examples/speech_synthesis/utils.py 9 | def batch_mel_cepstral_distortion(y1, y2, config): 10 | """ 11 | https://arxiv.org/pdf/2011.03568.pdf 12 | The root mean squared error computed on 13-dimensional MFCC using DTW for 13 | alignment. MFCC features are computed from an 80-channel log-mel 14 | spectrogram using a 50ms Hann window and hop of 12.5ms. 15 | y1: list of arrays of waveforms of type double 16 | y2: list of arrays of waveforms of type double 17 | sr: sampling rate 18 | """ 19 | 20 | return batch_compute_distortion( 21 | y1, 22 | y2, 23 | config.sr, 24 | lambda y: config.mfcc_fn.to(y1[0].device)(y).transpose(-1, -2), 25 | 'compute_rms_dist', 26 | config.norm_align_type, 27 | ) 28 | 29 | def main(gt_path, synth_path): 30 | 31 | # Load Audio 32 | x_gt, sr_gt = librosa.load(gt_path) 33 | x_synth, sr_synth = librosa.load(synth_path) 34 | 35 | # Load Config 36 | config = GlobalConfig() 37 | 38 | # Initiate Lists of Tensors for Batched DTW 39 | x1 = [] 40 | x2 = [] 41 | 42 | # Append Each Tensor 43 | x1.append(torch.Tensor(x_gt)) 44 | x2.append(torch.Tensor(x_synth)) 45 | 46 | return print(batch_mel_cepstral_distortion(x1, x2, config)) 47 | 48 | if __name__ == "__main__": 49 | 50 | parser = argparse.ArgumentParser(description='Script to calculate the Mel Cepstral Distortion between the ground truth and the synthesized audios.') 51 | parser.add_argument("--gt_path", required = True, type=str, help = 'Path to corresponding ground truth audio in .wav format') 52 | parser.add_argument("--synth_path", required = True, type=str, help = 'Path to corresponding synthesized audio in .wav format') 53 | 54 | args = parser.parse_args() 55 | 56 | gt_path = args.gt_path 57 | synth_path = args.synth_path 58 | 59 | main(gt_path, synth_path) -------------------------------------------------------------------------------- /metrics/MSD.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | from config.global_config import GlobalConfig 3 | import librosa 4 | import torchaudio 5 | import torch 6 | from metrics.helpers import batch_compute_distortion 7 | 8 | # https://github.com/pytorch/fairseq/blob/fcca32258c8e8bcc9f9890bf4714fa2f96b6b3e1/examples/speech_synthesis/utils.py 9 | def batch_mel_spectral_distortion( 10 | y1, y2, config 11 | ): 12 | """ 13 | https://arxiv.org/pdf/2011.03568.pdf 14 | Same as Mel Cepstral Distortion, but computed on log-mel spectrograms. 15 | """ 16 | offset = 1e-6 17 | return batch_compute_distortion( 18 | y1, y2, config.sr, lambda y: torch.log(config.mel_fn.to(y1[0].device)(y) + offset).transpose(-1, -2), 19 | 'compute_rms_dist', config.norm_align_type 20 | ) 21 | 22 | def main(gt_path, synth_path): 23 | 24 | # Load Audio 25 | x_gt, _ = librosa.load(gt_path) 26 | x_synth, _ = librosa.load(synth_path) 27 | 28 | # Load Config 29 | config = GlobalConfig() 30 | 31 | # Initiate Lists of Tensors for Batched DTW 32 | x1 = [] 33 | x2 = [] 34 | 35 | # Append Each Tensor 36 | x1.append(torch.Tensor(x_gt)) 37 | x2.append(torch.Tensor(x_synth)) 38 | 39 | return print(batch_mel_spectral_distortion(x1, x2, config)) 40 | 41 | if __name__ == "__main__": 42 | 43 | parser = argparse.ArgumentParser(description='Script to calculate the Mel Spectral Distortion between the ground truth and the synthesized audios.') 44 | parser.add_argument("--gt_path", required = True, type=str, help = 'Path to corresponding ground truth audio in .wav format') 45 | parser.add_argument("--synth_path", required = True, type=str, help = 'Path to corresponding synthesized audio in .wav format') 46 | 47 | args = parser.parse_args() 48 | 49 | gt_path = args.gt_path 50 | synth_path = args.synth_path 51 | 52 | main(gt_path, synth_path) -------------------------------------------------------------------------------- /metrics/VDE.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | from distutils.command.config import config 3 | import librosa 4 | import numpy as np 5 | from audio.pitch import yin, dio 6 | from config.global_config import GlobalConfig 7 | from metrics.helpers import same_t_in_true_and_est, voicing_decision_error_frames 8 | 9 | # https://github.com/bastibe/MAPS-Scripts/blob/master/helper.py 10 | @same_t_in_true_and_est 11 | def voicing_decision_error(true_t, true_f, est_t, est_f): 12 | """Percentages of frames that contain error in the voicing 13 | decision, according to the reference. 14 | """ 15 | vde_frames = voicing_decision_error_frames( 16 | true_t, true_f, est_t, est_f 17 | ) 18 | return np.sum(vde_frames) / (len(true_t)) 19 | 20 | def main(gt_path, synth_path, pitch_algorithm): 21 | 22 | # Load Audio 23 | x_gt, _ = librosa.load(gt_path) 24 | x_synth, _ = librosa.load(synth_path) 25 | 26 | # Load Config 27 | config = GlobalConfig() 28 | 29 | # Compute Pitch 30 | pitch_gt = eval(pitch_algorithm)(x_gt, config) 31 | pitch_synth = eval(pitch_algorithm)(x_synth, config) 32 | 33 | return print(voicing_decision_error(np.array(pitch_gt['times']), np.array(pitch_gt['pitches']), np.array(pitch_synth['times']), np.array(pitch_synth['pitches']))) 34 | 35 | if __name__ == "__main__": 36 | 37 | parser = argparse.ArgumentParser(description='Script to calculate the Voicing Decision Error between the ground truth and the synthesized audios.') 38 | parser.add_argument("--gt_path", required = True, type=str, help = 'Path to corresponding ground truth audio in .wav format') 39 | parser.add_argument("--synth_path", required = True, type=str, help = 'Path to corresponding synthesized audio in .wav format') 40 | parser.add_argument("--pitch_algorithm", required = True, type=str, choices = ["dio", "yin", "pyin"], help = 'Choose method of computing pitch') 41 | 42 | args = parser.parse_args() 43 | 44 | gt_path = args.gt_path 45 | synth_path = args.synth_path 46 | pitch_algorithm = args.pitch_algorithm 47 | 48 | main(gt_path, synth_path, pitch_algorithm) -------------------------------------------------------------------------------- /metrics/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AI-Unicamp/TTS-Objective-Metrics/e183fd0f4ec484173b13b86ae315c862af9db888/metrics/__init__.py -------------------------------------------------------------------------------- /metrics/dists.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | # try to make the distance function more general (only between two vectors and the expand it?) 4 | def compute_lp_dist(x1, x2, p): 5 | """compute an (m, n) Lp distance matrix from (m, d) and (n, d) matrices 6 | with all combination of distances between m and n""" 7 | return torch.cdist(x1.unsqueeze(0), x2.unsqueeze(0), p=p).squeeze(0) 8 | 9 | def compute_rms_dist(x1, x2): 10 | l2_dist = compute_lp_dist(x1, x2, 2) 11 | return (l2_dist.pow(2) / x1.size(1)).pow(0.5) -------------------------------------------------------------------------------- /metrics/helpers.py: -------------------------------------------------------------------------------- 1 | # https://github.com/bastibe/MAPS-Scripts/blob/master/helper.py 2 | # GPL 3.0 LICENSE 3 | import numpy as np 4 | import torch 5 | from metrics.DTW import batch_dynamic_time_warping 6 | from scipy.interpolate import interp1d 7 | 8 | def same_t_in_true_and_est(func): 9 | def new_func(true_t, true_f, est_t, est_f): 10 | assert type(true_t) is np.ndarray 11 | assert type(true_f) is np.ndarray 12 | assert type(est_t) is np.ndarray 13 | assert type(est_f) is np.ndarray 14 | 15 | interpolated_f = interp1d( 16 | est_t, est_f, bounds_error=False, kind='nearest', fill_value=0 17 | )(true_t) 18 | return func(true_t, true_f, true_t, interpolated_f) 19 | return new_func 20 | 21 | def gross_pitch_error_frames(true_t, true_f, est_t, est_f, eps=1e-8): 22 | voiced_frames = true_voiced_frames(true_t, true_f, est_t, est_f) 23 | true_f_p_eps = [x + eps for x in true_f] 24 | pitch_error_frames = np.abs(est_f / true_f_p_eps - 1) > 0.2 25 | return voiced_frames & pitch_error_frames 26 | 27 | def voicing_decision_error_frames(true_t, true_f, est_t, est_f): 28 | return (est_f != 0) != (true_f != 0) 29 | 30 | def true_voiced_frames(true_t, true_f, est_t, est_f): 31 | return (est_f != 0) & (true_f != 0) 32 | 33 | 34 | def batch_compute_distortion(y1, y2, sr, feat_fn, dist_fn, normalize_align_type): 35 | # Distances, Sizes, Spectrograms of Ref, Spectrograms of Synths 36 | s, x1, x2 = [], [], [] 37 | 38 | # Loop throug each (Ref, Synth) pair 39 | for cur_y1, cur_y2 in zip(y1, y2): 40 | assert cur_y1.ndim == 1 and cur_y2.ndim == 1 41 | 42 | # Compute Mel-Specs for MSD or MFCC for MCD 43 | cur_x1 = feat_fn(cur_y1) 44 | cur_x2 = feat_fn(cur_y2) 45 | x1.append(cur_x1) 46 | x2.append(cur_x2) 47 | 48 | # Save Sizes for UnPadding 49 | size = torch.empty(cur_x1.size(0), cur_x2.size(0)) 50 | s.append(size.size()) 51 | s = torch.LongTensor(s).to(cur_y1.device) 52 | 53 | # Get DTW D matrix, Minimal Cost Paths, Optimal Path 54 | return batch_dynamic_time_warping(x1,x2,dist_fn, normalize_align_type)['norm_align_costs'] 55 | 56 | 57 | def add_basic_stats(dic): 58 | mean = np.mean(list(dic.values())) 59 | std = np.std(list(dic.values())) 60 | max = np.max(list(dic.values())) 61 | argmax = float(np.argmax(list(dic.values()))) 62 | min = np.min(list(dic.values())) 63 | argmin = float(np.argmin(list(dic.values()))) 64 | dic['stats'] = {'mean':mean, 'std':std, 'max':max, 'argmax':argmax, 'min':min, 'argmin':argmin} 65 | return dic -------------------------------------------------------------------------------- /metrics/moments.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import librosa 3 | import numpy as np 4 | from audio.pitch import yin, dio 5 | from config.global_config import GlobalConfig 6 | from scipy.stats import tstd, skew, kurtosis 7 | 8 | def estimate_moments(pitch_gt, pitch_skew): 9 | gt_moments = {'std':tstd(pitch_gt), 'skew':skew(pitch_gt), 'kurt':kurtosis(pitch_gt)} 10 | synth_moments = {'std':tstd(pitch_skew), 'skew':skew(pitch_skew), 'kurt':kurtosis(pitch_skew)} 11 | return {'GT': gt_moments, 'SYNTH':synth_moments} 12 | 13 | def main(gt_path, synth_path, pitch_algorithm): 14 | 15 | # Load Audio 16 | x_gt, _ = librosa.load(gt_path) 17 | x_synth, _ = librosa.load(synth_path) 18 | 19 | # Load Config 20 | config = GlobalConfig() 21 | 22 | # Compute Pitch 23 | pitch_gt = eval(pitch_algorithm)(x_gt, config)['pitches'] 24 | pitch_synth = eval(pitch_algorithm)(x_synth, config)['pitches'] 25 | 26 | # Compute Moments 27 | return print(estimate_moments(pitch_gt, pitch_synth)) 28 | 29 | if __name__ == "__main__": 30 | 31 | parser = argparse.ArgumentParser(description='Script to estimate the Statistical Moments of the ground truth and the synthesized audios.') 32 | parser.add_argument("--gt_path", required = True, type=str, help = 'Path to corresponding ground truth audio in .wav format') 33 | parser.add_argument("--synth_path", required = True, type=str, help = 'Path to corresponding synthesized audio in .wav format') 34 | parser.add_argument("--pitch_algorithm", required = True, type=str, choices = ["dio", "yin"], help = 'Choose method of computing pitch') 35 | 36 | args = parser.parse_args() 37 | 38 | gt_path = args.gt_path 39 | synth_path = args.synth_path 40 | pitch_algorithm = args.pitch_algorithm 41 | 42 | main(gt_path, synth_path, pitch_algorithm) -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AI-Unicamp/TTS-Objective-Metrics/e183fd0f4ec484173b13b86ae315c862af9db888/requirements.txt --------------------------------------------------------------------------------