SayaSS commited on
Commit
6086416
1 Parent(s): 23a92a4
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. Eng_docs.md +109 -0
  2. LICENSE +407 -0
  3. README.md +4 -6
  4. app.py +75 -0
  5. cluster/__init__.py +29 -0
  6. cluster/__pycache__/__init__.cpython-38.pyc +0 -0
  7. cluster/train_cluster.py +89 -0
  8. configs/config.json +64 -0
  9. cvec/checkpoint_best_legacy_500.pt +3 -0
  10. data_utils.py +142 -0
  11. flask_api.py +56 -0
  12. hubert/__init__.py +0 -0
  13. hubert/__pycache__/__init__.cpython-38.pyc +0 -0
  14. hubert/__pycache__/hubert_model.cpython-38.pyc +0 -0
  15. hubert/checkpoint_best_legacy_500.pt +3 -0
  16. hubert/hubert_model.py +222 -0
  17. hubert/hubert_model_onnx.py +217 -0
  18. inference/__init__.py +0 -0
  19. inference/__pycache__/__init__.cpython-38.pyc +0 -0
  20. inference/__pycache__/infer_tool.cpython-38.pyc +0 -0
  21. inference/__pycache__/slicer.cpython-38.pyc +0 -0
  22. inference/chunks_temp.json +1 -0
  23. inference/infer_tool.py +233 -0
  24. inference/infer_tool_grad.py +160 -0
  25. inference/slicer.py +142 -0
  26. inference_main.py +100 -0
  27. models.py +420 -0
  28. models/tannhauser/config.json +93 -0
  29. models/tannhauser/tannhauser.pth +3 -0
  30. models/teio/config.json +93 -0
  31. models/teio/teio.pth +3 -0
  32. modules/__init__.py +0 -0
  33. modules/__pycache__/__init__.cpython-38.pyc +0 -0
  34. modules/__pycache__/attentions.cpython-38.pyc +0 -0
  35. modules/__pycache__/commons.cpython-38.pyc +0 -0
  36. modules/__pycache__/modules.cpython-38.pyc +0 -0
  37. modules/attentions.py +349 -0
  38. modules/commons.py +188 -0
  39. modules/ddsp.py +190 -0
  40. modules/losses.py +61 -0
  41. modules/mel_processing.py +112 -0
  42. modules/modules.py +342 -0
  43. onnx/model_onnx.py +328 -0
  44. onnx/model_onnx_48k.py +328 -0
  45. onnx/onnx_export.py +73 -0
  46. onnx/onnx_export_48k.py +73 -0
  47. preprocess_flist_config.py +67 -0
  48. preprocess_hubert_f0.py +62 -0
  49. requirements.txt +21 -0
  50. resample.py +48 -0
Eng_docs.md ADDED
@@ -0,0 +1,109 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SoftVC VITS Singing Voice Conversion
2
+
3
+ ## Updates
4
+ > According to incomplete statistics, it seems that training with multiple speakers may lead to **worsened leaking of voice timbre**. It is not recommended to train models with more than 5 speakers. The current suggestion is to try to train models with only a single speaker if you want to achieve a voice timbre that is more similar to the target.
5
+ > Fixed the issue with unwanted staccato, improving audio quality by a decent amount.\
6
+ > The 2.0 version has been moved to the 2.0 branch.\
7
+ > Version 3.0 uses the code structure of FreeVC, which isn't compatible with older versions.\
8
+ > Compared to [DiffSVC](https://github.com/prophesier/diff-svc) , diffsvc performs much better when the training data is of extremely high quality, but this repository may perform better on datasets with lower quality. Additionally, this repository is much faster in terms of inference speed compared to diffsvc.
9
+
10
+ ## Model Overview
11
+ A singing voice coversion (SVC) model, using the SoftVC encoder to extract features from the input audio, sent into VITS along with the F0 to replace the original input to acheive a voice conversion effect. Additionally, changing the vocoder to [NSF HiFiGAN](https://github.com/openvpi/DiffSinger/tree/refactor/modules/nsf_hifigan) to fix the issue with unwanted staccato.
12
+
13
+ ## Notice
14
+ + The current branch is the 32kHz version, which requires less vram during inferencing, as well as faster inferencing speeds, and datasets for said branch take up less disk space. Thus the 32 kHz branch is recommended for use.
15
+ + If you want to train 48 kHz variant models, switch to the [main branch](https://github.com/innnky/so-vits-svc/tree/main).
16
+
17
+
18
+ ## Required models
19
+ + soft vc hubert:[hubert-soft-0d54a1f4.pt](https://github.com/bshall/hubert/releases/download/v0.1/hubert-soft-0d54a1f4.pt)
20
+ + Place under `hubert`.
21
+ + Pretrained models [G_0.pth](https://huggingface.co/innnky/sovits_pretrained/resolve/main/G_0.pth) and [D_0.pth](https://huggingface.co/innnky/sovits_pretrained/resolve/main/D_0.pth)
22
+ + Place under `logs/32k`.
23
+ + Pretrained models are required, because from experiments, training from scratch can be rather unpredictable to say the least, and training with a pretrained model can greatly improve training speeds.
24
+ + The pretrained model includes云灏, 即霜, 辉宇·星AI, 派蒙, and 绫地宁宁, covering the common ranges of both male and female voices, and so it can be seen as a rather universal pretrained model.
25
+ + The pretrained model exludes the `optimizer speaker_embedding` section, rendering it only usable for pretraining and incapable of inferencing with.
26
+ ```shell
27
+ # For simple downloading.
28
+ # hubert
29
+ wget -P hubert/ https://github.com/bshall/hubert/releases/download/v0.1/hubert-soft-0d54a1f4.pt
30
+ # G&D pretrained models
31
+ wget -P logs/32k/ https://huggingface.co/innnky/sovits_pretrained/resolve/main/G_0.pth
32
+ wget -P logs/32k/ https://huggingface.co/innnky/sovits_pretrained/resolve/main/D_0.pth
33
+
34
+ ```
35
+
36
+ ## Colab notebook script for dataset creation and training.
37
+ [colab training notebook](https://colab.research.google.com/drive/1rCUOOVG7-XQlVZuWRAj5IpGrMM8t07pE?usp=sharing)
38
+
39
+ ## Dataset preparation
40
+ All that is required is that the data be put under the `dataset_raw` folder in the structure format provided below.
41
+ ```shell
42
+ dataset_raw
43
+ ├───speaker0
44
+ │ ├───xxx1-xxx1.wav
45
+ │ ├───...
46
+ │ └───Lxx-0xx8.wav
47
+ └───speaker1
48
+ ├───xx2-0xxx2.wav
49
+ ├───...
50
+ └───xxx7-xxx007.wav
51
+ ```
52
+
53
+ ## Data pre-processing.
54
+ 1. Resample to 32khz
55
+
56
+ ```shell
57
+ python resample.py
58
+ ```
59
+ 2. Automatically sort out training set, validation set, test set, and automatically generate configuration files.
60
+ ```shell
61
+ python preprocess_flist_config.py
62
+ # Notice.
63
+ # The n_speakers value in the config will be set automatically according to the amount of speakers in the dataset.
64
+ # To reserve space for additionally added speakers in the dataset, the n_speakers value will be be set to twice the actual amount.
65
+ # If you want even more space for adding more data, you can edit the n_speakers value in the config after runing this step.
66
+ # This can not be changed after training starts.
67
+ ```
68
+ 3. Generate hubert and F0 features/
69
+ ```shell
70
+ python preprocess_hubert_f0.py
71
+ ```
72
+ After running the step above, the `dataset` folder will contain all the pre-processed data, you can delete the `dataset_raw` folder after that.
73
+
74
+ ## Training.
75
+ ```shell
76
+ python train.py -c configs/config.json -m 32k
77
+ ```
78
+
79
+ ## Inferencing.
80
+
81
+ Use [inference_main.py](inference_main.py)
82
+ + Edit `model_path` to your newest checkpoint.
83
+ + Place the input audio under the `raw` folder.
84
+ + Change `clean_names` to the output file name.
85
+ + Use `trans` to edit the pitch shifting amount (semitones).
86
+ + Change `spk_list` to the speaker name.
87
+
88
+ ## Onnx Exporting.
89
+ ### **When exporting Onnx, please make sure you re-clone the whole repository!!!**
90
+ Use [onnx_export.py](onnx_export.py)
91
+ + Create a new folder called `checkpoints`.
92
+ + Create a project folder in `checkpoints` folder with the desired name for your project, let's use `myproject` as example. Folder structure looks like `./checkpoints/myproject`.
93
+ + Rename your model to `model.pth`, rename your config file to `config.json` then move them into `myproject` folder.
94
+ + Modify [onnx_export.py](onnx_export.py) where `path = "NyaruTaffy"`, change `NyaruTaffy` to your project name, here it will be `path = "myproject"`.
95
+ + Run [onnx_export.py](onnx_export.py)
96
+ + Once it finished, a `model.onnx` will be generated in `myproject` folder, that's the model you just exported.
97
+ + Notice: if you want to export a 48K model, please follow the instruction below or use `model_onnx_48k.py` directly.
98
+ + Open [model_onnx.py](model_onnx.py) and change `hps={"sampling_rate": 32000...}` to `hps={"sampling_rate": 48000}` in class `SynthesizerTrn`.
99
+ + Open [nvSTFT](/vdecoder/hifigan/nvSTFT.py) and replace all `32000` with `48000`
100
+ ### Onnx Model UI Support
101
+ + [MoeSS](https://github.com/NaruseMioShirakana/MoeSS)
102
+ + All training function and transformation are removed, only if they are all removed you are actually using Onnx.
103
+
104
+ ## Gradio (WebUI)
105
+ Use [sovits_gradio.py](sovits_gradio.py) to run Gradio WebUI
106
+ + Create a new folder called `checkpoints`.
107
+ + Create a project folder in `checkpoints` folder with the desired name for your project, let's use `myproject` as example. Folder structure looks like `./checkpoints/myproject`.
108
+ + Rename your model to `model.pth`, rename your config file to `config.json` then move them into `myproject` folder.
109
+ + Run [sovits_gradio.py](sovits_gradio.py)
LICENSE ADDED
@@ -0,0 +1,407 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Attribution-NonCommercial 4.0 International
2
+
3
+ =======================================================================
4
+
5
+ Creative Commons Corporation ("Creative Commons") is not a law firm and
6
+ does not provide legal services or legal advice. Distribution of
7
+ Creative Commons public licenses does not create a lawyer-client or
8
+ other relationship. Creative Commons makes its licenses and related
9
+ information available on an "as-is" basis. Creative Commons gives no
10
+ warranties regarding its licenses, any material licensed under their
11
+ terms and conditions, or any related information. Creative Commons
12
+ disclaims all liability for damages resulting from their use to the
13
+ fullest extent possible.
14
+
15
+ Using Creative Commons Public Licenses
16
+
17
+ Creative Commons public licenses provide a standard set of terms and
18
+ conditions that creators and other rights holders may use to share
19
+ original works of authorship and other material subject to copyright
20
+ and certain other rights specified in the public license below. The
21
+ following considerations are for informational purposes only, are not
22
+ exhaustive, and do not form part of our licenses.
23
+
24
+ Considerations for licensors: Our public licenses are
25
+ intended for use by those authorized to give the public
26
+ permission to use material in ways otherwise restricted by
27
+ copyright and certain other rights. Our licenses are
28
+ irrevocable. Licensors should read and understand the terms
29
+ and conditions of the license they choose before applying it.
30
+ Licensors should also secure all rights necessary before
31
+ applying our licenses so that the public can reuse the
32
+ material as expected. Licensors should clearly mark any
33
+ material not subject to the license. This includes other CC-
34
+ licensed material, or material used under an exception or
35
+ limitation to copyright. More considerations for licensors:
36
+ wiki.creativecommons.org/Considerations_for_licensors
37
+
38
+ Considerations for the public: By using one of our public
39
+ licenses, a licensor grants the public permission to use the
40
+ licensed material under specified terms and conditions. If
41
+ the licensor's permission is not necessary for any reason--for
42
+ example, because of any applicable exception or limitation to
43
+ copyright--then that use is not regulated by the license. Our
44
+ licenses grant only permissions under copyright and certain
45
+ other rights that a licensor has authority to grant. Use of
46
+ the licensed material may still be restricted for other
47
+ reasons, including because others have copyright or other
48
+ rights in the material. A licensor may make special requests,
49
+ such as asking that all changes be marked or described.
50
+ Although not required by our licenses, you are encouraged to
51
+ respect those requests where reasonable. More considerations
52
+ for the public:
53
+ wiki.creativecommons.org/Considerations_for_licensees
54
+
55
+ =======================================================================
56
+
57
+ Creative Commons Attribution-NonCommercial 4.0 International Public
58
+ License
59
+
60
+ By exercising the Licensed Rights (defined below), You accept and agree
61
+ to be bound by the terms and conditions of this Creative Commons
62
+ Attribution-NonCommercial 4.0 International Public License ("Public
63
+ License"). To the extent this Public License may be interpreted as a
64
+ contract, You are granted the Licensed Rights in consideration of Your
65
+ acceptance of these terms and conditions, and the Licensor grants You
66
+ such rights in consideration of benefits the Licensor receives from
67
+ making the Licensed Material available under these terms and
68
+ conditions.
69
+
70
+
71
+ Section 1 -- Definitions.
72
+
73
+ a. Adapted Material means material subject to Copyright and Similar
74
+ Rights that is derived from or based upon the Licensed Material
75
+ and in which the Licensed Material is translated, altered,
76
+ arranged, transformed, or otherwise modified in a manner requiring
77
+ permission under the Copyright and Similar Rights held by the
78
+ Licensor. For purposes of this Public License, where the Licensed
79
+ Material is a musical work, performance, or sound recording,
80
+ Adapted Material is always produced where the Licensed Material is
81
+ synched in timed relation with a moving image.
82
+
83
+ b. Adapter's License means the license You apply to Your Copyright
84
+ and Similar Rights in Your contributions to Adapted Material in
85
+ accordance with the terms and conditions of this Public License.
86
+
87
+ c. Copyright and Similar Rights means copyright and/or similar rights
88
+ closely related to copyright including, without limitation,
89
+ performance, broadcast, sound recording, and Sui Generis Database
90
+ Rights, without regard to how the rights are labeled or
91
+ categorized. For purposes of this Public License, the rights
92
+ specified in Section 2(b)(1)-(2) are not Copyright and Similar
93
+ Rights.
94
+ d. Effective Technological Measures means those measures that, in the
95
+ absence of proper authority, may not be circumvented under laws
96
+ fulfilling obligations under Article 11 of the WIPO Copyright
97
+ Treaty adopted on December 20, 1996, and/or similar international
98
+ agreements.
99
+
100
+ e. Exceptions and Limitations means fair use, fair dealing, and/or
101
+ any other exception or limitation to Copyright and Similar Rights
102
+ that applies to Your use of the Licensed Material.
103
+
104
+ f. Licensed Material means the artistic or literary work, database,
105
+ or other material to which the Licensor applied this Public
106
+ License.
107
+
108
+ g. Licensed Rights means the rights granted to You subject to the
109
+ terms and conditions of this Public License, which are limited to
110
+ all Copyright and Similar Rights that apply to Your use of the
111
+ Licensed Material and that the Licensor has authority to license.
112
+
113
+ h. Licensor means the individual(s) or entity(ies) granting rights
114
+ under this Public License.
115
+
116
+ i. NonCommercial means not primarily intended for or directed towards
117
+ commercial advantage or monetary compensation. For purposes of
118
+ this Public License, the exchange of the Licensed Material for
119
+ other material subject to Copyright and Similar Rights by digital
120
+ file-sharing or similar means is NonCommercial provided there is
121
+ no payment of monetary compensation in connection with the
122
+ exchange.
123
+
124
+ j. Share means to provide material to the public by any means or
125
+ process that requires permission under the Licensed Rights, such
126
+ as reproduction, public display, public performance, distribution,
127
+ dissemination, communication, or importation, and to make material
128
+ available to the public including in ways that members of the
129
+ public may access the material from a place and at a time
130
+ individually chosen by them.
131
+
132
+ k. Sui Generis Database Rights means rights other than copyright
133
+ resulting from Directive 96/9/EC of the European Parliament and of
134
+ the Council of 11 March 1996 on the legal protection of databases,
135
+ as amended and/or succeeded, as well as other essentially
136
+ equivalent rights anywhere in the world.
137
+
138
+ l. You means the individual or entity exercising the Licensed Rights
139
+ under this Public License. Your has a corresponding meaning.
140
+
141
+
142
+ Section 2 -- Scope.
143
+
144
+ a. License grant.
145
+
146
+ 1. Subject to the terms and conditions of this Public License,
147
+ the Licensor hereby grants You a worldwide, royalty-free,
148
+ non-sublicensable, non-exclusive, irrevocable license to
149
+ exercise the Licensed Rights in the Licensed Material to:
150
+
151
+ a. reproduce and Share the Licensed Material, in whole or
152
+ in part, for NonCommercial purposes only; and
153
+
154
+ b. produce, reproduce, and Share Adapted Material for
155
+ NonCommercial purposes only.
156
+
157
+ 2. Exceptions and Limitations. For the avoidance of doubt, where
158
+ Exceptions and Limitations apply to Your use, this Public
159
+ License does not apply, and You do not need to comply with
160
+ its terms and conditions.
161
+
162
+ 3. Term. The term of this Public License is specified in Section
163
+ 6(a).
164
+
165
+ 4. Media and formats; technical modifications allowed. The
166
+ Licensor authorizes You to exercise the Licensed Rights in
167
+ all media and formats whether now known or hereafter created,
168
+ and to make technical modifications necessary to do so. The
169
+ Licensor waives and/or agrees not to assert any right or
170
+ authority to forbid You from making technical modifications
171
+ necessary to exercise the Licensed Rights, including
172
+ technical modifications necessary to circumvent Effective
173
+ Technological Measures. For purposes of this Public License,
174
+ simply making modifications authorized by this Section 2(a)
175
+ (4) never produces Adapted Material.
176
+
177
+ 5. Downstream recipients.
178
+
179
+ a. Offer from the Licensor -- Licensed Material. Every
180
+ recipient of the Licensed Material automatically
181
+ receives an offer from the Licensor to exercise the
182
+ Licensed Rights under the terms and conditions of this
183
+ Public License.
184
+
185
+ b. No downstream restrictions. You may not offer or impose
186
+ any additional or different terms or conditions on, or
187
+ apply any Effective Technological Measures to, the
188
+ Licensed Material if doing so restricts exercise of the
189
+ Licensed Rights by any recipient of the Licensed
190
+ Material.
191
+
192
+ 6. No endorsement. Nothing in this Public License constitutes or
193
+ may be construed as permission to assert or imply that You
194
+ are, or that Your use of the Licensed Material is, connected
195
+ with, or sponsored, endorsed, or granted official status by,
196
+ the Licensor or others designated to receive attribution as
197
+ provided in Section 3(a)(1)(A)(i).
198
+
199
+ b. Other rights.
200
+
201
+ 1. Moral rights, such as the right of integrity, are not
202
+ licensed under this Public License, nor are publicity,
203
+ privacy, and/or other similar personality rights; however, to
204
+ the extent possible, the Licensor waives and/or agrees not to
205
+ assert any such rights held by the Licensor to the limited
206
+ extent necessary to allow You to exercise the Licensed
207
+ Rights, but not otherwise.
208
+
209
+ 2. Patent and trademark rights are not licensed under this
210
+ Public License.
211
+
212
+ 3. To the extent possible, the Licensor waives any right to
213
+ collect royalties from You for the exercise of the Licensed
214
+ Rights, whether directly or through a collecting society
215
+ under any voluntary or waivable statutory or compulsory
216
+ licensing scheme. In all other cases the Licensor expressly
217
+ reserves any right to collect such royalties, including when
218
+ the Licensed Material is used other than for NonCommercial
219
+ purposes.
220
+
221
+
222
+ Section 3 -- License Conditions.
223
+
224
+ Your exercise of the Licensed Rights is expressly made subject to the
225
+ following conditions.
226
+
227
+ a. Attribution.
228
+
229
+ 1. If You Share the Licensed Material (including in modified
230
+ form), You must:
231
+
232
+ a. retain the following if it is supplied by the Licensor
233
+ with the Licensed Material:
234
+
235
+ i. identification of the creator(s) of the Licensed
236
+ Material and any others designated to receive
237
+ attribution, in any reasonable manner requested by
238
+ the Licensor (including by pseudonym if
239
+ designated);
240
+
241
+ ii. a copyright notice;
242
+
243
+ iii. a notice that refers to this Public License;
244
+
245
+ iv. a notice that refers to the disclaimer of
246
+ warranties;
247
+
248
+ v. a URI or hyperlink to the Licensed Material to the
249
+ extent reasonably practicable;
250
+
251
+ b. indicate if You modified the Licensed Material and
252
+ retain an indication of any previous modifications; and
253
+
254
+ c. indicate the Licensed Material is licensed under this
255
+ Public License, and include the text of, or the URI or
256
+ hyperlink to, this Public License.
257
+
258
+ 2. You may satisfy the conditions in Section 3(a)(1) in any
259
+ reasonable manner based on the medium, means, and context in
260
+ which You Share the Licensed Material. For example, it may be
261
+ reasonable to satisfy the conditions by providing a URI or
262
+ hyperlink to a resource that includes the required
263
+ information.
264
+
265
+ 3. If requested by the Licensor, You must remove any of the
266
+ information required by Section 3(a)(1)(A) to the extent
267
+ reasonably practicable.
268
+
269
+ 4. If You Share Adapted Material You produce, the Adapter's
270
+ License You apply must not prevent recipients of the Adapted
271
+ Material from complying with this Public License.
272
+
273
+
274
+ Section 4 -- Sui Generis Database Rights.
275
+
276
+ Where the Licensed Rights include Sui Generis Database Rights that
277
+ apply to Your use of the Licensed Material:
278
+
279
+ a. for the avoidance of doubt, Section 2(a)(1) grants You the right
280
+ to extract, reuse, reproduce, and Share all or a substantial
281
+ portion of the contents of the database for NonCommercial purposes
282
+ only;
283
+
284
+ b. if You include all or a substantial portion of the database
285
+ contents in a database in which You have Sui Generis Database
286
+ Rights, then the database in which You have Sui Generis Database
287
+ Rights (but not its individual contents) is Adapted Material; and
288
+
289
+ c. You must comply with the conditions in Section 3(a) if You Share
290
+ all or a substantial portion of the contents of the database.
291
+
292
+ For the avoidance of doubt, this Section 4 supplements and does not
293
+ replace Your obligations under this Public License where the Licensed
294
+ Rights include other Copyright and Similar Rights.
295
+
296
+
297
+ Section 5 -- Disclaimer of Warranties and Limitation of Liability.
298
+
299
+ a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
300
+ EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
301
+ AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
302
+ ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
303
+ IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
304
+ WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
305
+ PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
306
+ ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
307
+ KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
308
+ ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
309
+
310
+ b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
311
+ TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
312
+ NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
313
+ INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
314
+ COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
315
+ USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
316
+ ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
317
+ DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
318
+ IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
319
+
320
+ c. The disclaimer of warranties and limitation of liability provided
321
+ above shall be interpreted in a manner that, to the extent
322
+ possible, most closely approximates an absolute disclaimer and
323
+ waiver of all liability.
324
+
325
+
326
+ Section 6 -- Term and Termination.
327
+
328
+ a. This Public License applies for the term of the Copyright and
329
+ Similar Rights licensed here. However, if You fail to comply with
330
+ this Public License, then Your rights under this Public License
331
+ terminate automatically.
332
+
333
+ b. Where Your right to use the Licensed Material has terminated under
334
+ Section 6(a), it reinstates:
335
+
336
+ 1. automatically as of the date the violation is cured, provided
337
+ it is cured within 30 days of Your discovery of the
338
+ violation; or
339
+
340
+ 2. upon express reinstatement by the Licensor.
341
+
342
+ For the avoidance of doubt, this Section 6(b) does not affect any
343
+ right the Licensor may have to seek remedies for Your violations
344
+ of this Public License.
345
+
346
+ c. For the avoidance of doubt, the Licensor may also offer the
347
+ Licensed Material under separate terms or conditions or stop
348
+ distributing the Licensed Material at any time; however, doing so
349
+ will not terminate this Public License.
350
+
351
+ d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
352
+ License.
353
+
354
+
355
+ Section 7 -- Other Terms and Conditions.
356
+
357
+ a. The Licensor shall not be bound by any additional or different
358
+ terms or conditions communicated by You unless expressly agreed.
359
+
360
+ b. Any arrangements, understandings, or agreements regarding the
361
+ Licensed Material not stated herein are separate from and
362
+ independent of the terms and conditions of this Public License.
363
+
364
+
365
+ Section 8 -- Interpretation.
366
+
367
+ a. For the avoidance of doubt, this Public License does not, and
368
+ shall not be interpreted to, reduce, limit, restrict, or impose
369
+ conditions on any use of the Licensed Material that could lawfully
370
+ be made without permission under this Public License.
371
+
372
+ b. To the extent possible, if any provision of this Public License is
373
+ deemed unenforceable, it shall be automatically reformed to the
374
+ minimum extent necessary to make it enforceable. If the provision
375
+ cannot be reformed, it shall be severed from this Public License
376
+ without affecting the enforceability of the remaining terms and
377
+ conditions.
378
+
379
+ c. No term or condition of this Public License will be waived and no
380
+ failure to comply consented to unless expressly agreed to by the
381
+ Licensor.
382
+
383
+ d. Nothing in this Public License constitutes or may be interpreted
384
+ as a limitation upon, or waiver of, any privileges and immunities
385
+ that apply to the Licensor or You, including from the legal
386
+ processes of any jurisdiction or authority.
387
+
388
+ =======================================================================
389
+
390
+ Creative Commons is not a party to its public
391
+ licenses. Notwithstanding, Creative Commons may elect to apply one of
392
+ its public licenses to material it publishes and in those instances
393
+ will be considered the “Licensor.” The text of the Creative Commons
394
+ public licenses is dedicated to the public domain under the CC0 Public
395
+ Domain Dedication. Except for the limited purpose of indicating that
396
+ material is shared under a Creative Commons public license or as
397
+ otherwise permitted by the Creative Commons policies published at
398
+ creativecommons.org/policies, Creative Commons does not authorize the
399
+ use of the trademark "Creative Commons" or any other trademark or logo
400
+ of Creative Commons without its prior written consent including,
401
+ without limitation, in connection with any unauthorized modifications
402
+ to any of its public licenses or any other arrangements,
403
+ understandings, or agreements concerning use of licensed material. For
404
+ the avoidance of doubt, this paragraph does not form part of the
405
+ public licenses.
406
+
407
+ Creative Commons may be contacted at creativecommons.org.
README.md CHANGED
@@ -1,13 +1,11 @@
1
  ---
2
  title: Sovits Umamusume
3
- emoji: 💻
4
- colorFrom: blue
5
- colorTo: purple
6
  sdk: gradio
7
- sdk_version: 3.19.1
8
  app_file: app.py
9
  pinned: false
10
  license: mit
11
  ---
12
-
13
- Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
1
  ---
2
  title: Sovits Umamusume
3
+ emoji: 🐎
4
+ colorFrom: gray
5
+ colorTo: pink
6
  sdk: gradio
7
+ sdk_version: 3.18.0
8
  app_file: app.py
9
  pinned: false
10
  license: mit
11
  ---
 
 
app.py ADDED
@@ -0,0 +1,75 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import io
2
+ import os
3
+ import gradio as gr
4
+ import librosa
5
+ import numpy as np
6
+ import soundfile
7
+ from inference.infer_tool import Svc
8
+ import logging
9
+
10
+ logging.getLogger('numba').setLevel(logging.WARNING)
11
+ logging.getLogger('markdown_it').setLevel(logging.WARNING)
12
+ logging.getLogger('urllib3').setLevel(logging.WARNING)
13
+ logging.getLogger('matplotlib').setLevel(logging.WARNING)
14
+
15
+ def create_vc_fn(model, sid):
16
+ def vc_fn(input_audio, vc_transform, auto_f0):
17
+ if input_audio is None:
18
+ return "You need to upload an audio", None
19
+ sampling_rate, audio = input_audio
20
+ # print(audio.shape,sampling_rate)
21
+ duration = audio.shape[0] / sampling_rate
22
+ if duration > 45:
23
+ return "Please upload an audio file that is less than 45 seconds. If you need to generate a longer audio file, please use Colab.", None
24
+ audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
25
+ if len(audio.shape) > 1:
26
+ audio = librosa.to_mono(audio.transpose(1, 0))
27
+ if sampling_rate != 16000:
28
+ audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
29
+ out_wav_path = "temp.wav"
30
+ soundfile.write(out_wav_path, audio, 16000, format="wav")
31
+ out_audio, out_sr = model.infer(sid, vc_transform, out_wav_path,
32
+ auto_predict_f0=auto_f0,
33
+ )
34
+ return "Success", (44100, out_audio.cpu().numpy())
35
+ return vc_fn
36
+
37
+ if __name__ == '__main__':
38
+ models = []
39
+ for f in os.listdir("models"):
40
+ name = f
41
+ model = Svc(fr"models/{f}/{f}.pth", f"models/{f}/config.json")
42
+ cover = f"models/{f}/cover.png" if os.path.exists(f"models/{f}/cover.png") else None
43
+ models.append((name, cover, create_vc_fn(model, name)))
44
+ with gr.Blocks() as app:
45
+ gr.Markdown(
46
+ "# <center> Sovits Umamusume\n"
47
+ "## <center> The input audio should be clean and pure voice without background music.\n"
48
+ "![visitor badge](https://visitor-badge.glitch.me/badge?page_id=sayashi.vits-models)\n\n"
49
+ "[Open In Colab]()"
50
+ "\n\n"
51
+ "[Original Repo](https://github.com/innnky/so-vits-svc/tree/4.0)"
52
+ )
53
+ with gr.Tabs():
54
+ for (name, cover, vc_fn) in models:
55
+ with gr.TabItem(name):
56
+ with gr.Row():
57
+ gr.Markdown(
58
+ '<div align="center">'
59
+ f'<img style="width:auto;height:300px;" src="file/{cover}">' if cover else ""
60
+ '</div>'
61
+ )
62
+ with gr.Row():
63
+ with gr.Column():
64
+ vc_input = gr.Audio(label="Input audio (less than 45 seconds)")
65
+ vc_transform = gr.Number(label="vc_transform", value=0)
66
+ auto_f0 = gr.Checkbox(label="auto_f0", value=False)
67
+ vc_submit = gr.Button("Generate", variant="primary")
68
+ with gr.Column():
69
+ vc_output1 = gr.Textbox(label="Output Message")
70
+ vc_output2 = gr.Audio(label="Output Audio")
71
+ vc_submit.click(vc_fn, [vc_input, vc_transform, auto_f0], [vc_output1, vc_output2])
72
+ app.queue(concurrency_count=1).launch()
73
+
74
+
75
+
cluster/__init__.py ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import torch
3
+ from sklearn.cluster import KMeans
4
+
5
+ def get_cluster_model(ckpt_path):
6
+ checkpoint = torch.load(ckpt_path)
7
+ kmeans_dict = {}
8
+ for spk, ckpt in checkpoint.items():
9
+ km = KMeans(ckpt["n_features_in_"])
10
+ km.__dict__["n_features_in_"] = ckpt["n_features_in_"]
11
+ km.__dict__["_n_threads"] = ckpt["_n_threads"]
12
+ km.__dict__["cluster_centers_"] = ckpt["cluster_centers_"]
13
+ kmeans_dict[spk] = km
14
+ return kmeans_dict
15
+
16
+ def get_cluster_result(model, x, speaker):
17
+ """
18
+ x: np.array [t, 256]
19
+ return cluster class result
20
+ """
21
+ return model[speaker].predict(x)
22
+
23
+ def get_cluster_center_result(model, x,speaker):
24
+ """x: np.array [t, 256]"""
25
+ predict = model[speaker].predict(x)
26
+ return model[speaker].cluster_centers_[predict]
27
+
28
+ def get_center(model, x,speaker):
29
+ return model[speaker].cluster_centers_[x]
cluster/__pycache__/__init__.cpython-38.pyc ADDED
Binary file (1.09 kB). View file
 
cluster/train_cluster.py ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ from glob import glob
3
+ from pathlib import Path
4
+ import torch
5
+ import logging
6
+ import argparse
7
+ import torch
8
+ import numpy as np
9
+ from sklearn.cluster import KMeans, MiniBatchKMeans
10
+ import tqdm
11
+ logging.basicConfig(level=logging.INFO)
12
+ logger = logging.getLogger(__name__)
13
+ import time
14
+ import random
15
+
16
+ def train_cluster(in_dir, n_clusters, use_minibatch=True, verbose=False):
17
+
18
+ logger.info(f"Loading features from {in_dir}")
19
+ features = []
20
+ nums = 0
21
+ for path in tqdm.tqdm(in_dir.glob("*.soft.pt")):
22
+ features.append(torch.load(path).squeeze(0).numpy().T)
23
+ # print(features[-1].shape)
24
+ features = np.concatenate(features, axis=0)
25
+ print(nums, features.nbytes/ 1024**2, "MB , shape:",features.shape, features.dtype)
26
+ features = features.astype(np.float32)
27
+ logger.info(f"Clustering features of shape: {features.shape}")
28
+ t = time.time()
29
+ if use_minibatch:
30
+ kmeans = MiniBatchKMeans(n_clusters=n_clusters,verbose=verbose, batch_size=4096, max_iter=80).fit(features)
31
+ else:
32
+ kmeans = KMeans(n_clusters=n_clusters,verbose=verbose).fit(features)
33
+ print(time.time()-t, "s")
34
+
35
+ x = {
36
+ "n_features_in_": kmeans.n_features_in_,
37
+ "_n_threads": kmeans._n_threads,
38
+ "cluster_centers_": kmeans.cluster_centers_,
39
+ }
40
+ print("end")
41
+
42
+ return x
43
+
44
+
45
+ if __name__ == "__main__":
46
+
47
+ parser = argparse.ArgumentParser()
48
+ parser.add_argument('--dataset', type=Path, default="./dataset/44k",
49
+ help='path of training data directory')
50
+ parser.add_argument('--output', type=Path, default="logs/44k",
51
+ help='path of model output directory')
52
+
53
+ args = parser.parse_args()
54
+
55
+ checkpoint_dir = args.output
56
+ dataset = args.dataset
57
+ n_clusters = 10000
58
+
59
+ ckpt = {}
60
+ for spk in os.listdir(dataset):
61
+ if os.path.isdir(dataset/spk):
62
+ print(f"train kmeans for {spk}...")
63
+ in_dir = dataset/spk
64
+ x = train_cluster(in_dir, n_clusters, verbose=False)
65
+ ckpt[spk] = x
66
+
67
+ checkpoint_path = checkpoint_dir / f"kmeans_{n_clusters}.pt"
68
+ checkpoint_path.parent.mkdir(exist_ok=True, parents=True)
69
+ torch.save(
70
+ ckpt,
71
+ checkpoint_path,
72
+ )
73
+
74
+
75
+ # import cluster
76
+ # for spk in tqdm.tqdm(os.listdir("dataset")):
77
+ # if os.path.isdir(f"dataset/{spk}"):
78
+ # print(f"start kmeans inference for {spk}...")
79
+ # for feature_path in tqdm.tqdm(glob(f"dataset/{spk}/*.discrete.npy", recursive=True)):
80
+ # mel_path = feature_path.replace(".discrete.npy",".mel.npy")
81
+ # mel_spectrogram = np.load(mel_path)
82
+ # feature_len = mel_spectrogram.shape[-1]
83
+ # c = np.load(feature_path)
84
+ # c = utils.tools.repeat_expand_2d(torch.FloatTensor(c), feature_len).numpy()
85
+ # feature = c.T
86
+ # feature_class = cluster.get_cluster_result(feature, spk)
87
+ # np.save(feature_path.replace(".discrete.npy", ".discrete_class.npy"), feature_class)
88
+
89
+
configs/config.json ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "train": {
3
+ "log_interval": 200,
4
+ "eval_interval": 800,
5
+ "seed": 1234,
6
+ "epochs": 10000,
7
+ "learning_rate": 0.0001,
8
+ "betas": [
9
+ 0.8,
10
+ 0.99
11
+ ],
12
+ "eps": 1e-09,
13
+ "batch_size": 6,
14
+ "fp16_run": false,
15
+ "lr_decay": 0.999875,
16
+ "segment_size": 10240,
17
+ "init_lr_ratio": 1,
18
+ "warmup_epochs": 0,
19
+ "c_mel": 45,
20
+ "c_kl": 1.0,
21
+ "use_sr": true,
22
+ "max_speclen": 512,
23
+ "port": "8001"
24
+ },
25
+ "data": {
26
+ "training_files": "filelists/train.txt",
27
+ "validation_files": "filelists/val.txt",
28
+ "max_wav_value": 32768.0,
29
+ "sampling_rate": 44100,
30
+ "filter_length": 2048,
31
+ "hop_length": 512,
32
+ "win_length": 2048,
33
+ "n_mel_channels": 80,
34
+ "mel_fmin": 0.0,
35
+ "mel_fmax": 22050
36
+ },
37
+ "model": {
38
+ "inter_channels": 192,
39
+ "hidden_channels": 192,
40
+ "filter_channels": 768,
41
+ "n_heads": 2,
42
+ "n_layers": 6,
43
+ "kernel_size": 3,
44
+ "p_dropout": 0.1,
45
+ "resblock": "1",
46
+ "resblock_kernel_sizes": [3,7,11],
47
+ "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
48
+ "upsample_rates": [ 8, 8, 2, 2, 2],
49
+ "upsample_initial_channel": 512,
50
+ "upsample_kernel_sizes": [16,16, 4, 4, 4],
51
+ "n_layers_q": 3,
52
+ "use_spectral_norm": false,
53
+ "gin_channels": 256,
54
+ "ssl_dim": 256,
55
+ "n_speakers": 200
56
+ },
57
+ "spk": {
58
+ "jishuang": 0,
59
+ "huiyu": 1,
60
+ "nen": 2,
61
+ "paimon": 3,
62
+ "yunhao": 4
63
+ }
64
+ }
cvec/checkpoint_best_legacy_500.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:294a2e8c98136070a999e040ec98dfa5a99b88a7938181c56cc2ab0e2f6ce0e8
3
+ size 48501067
data_utils.py ADDED
@@ -0,0 +1,142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import time
2
+ import os
3
+ import random
4
+ import numpy as np
5
+ import torch
6
+ import torch.utils.data
7
+
8
+ import modules.commons as commons
9
+ import utils
10
+ from modules.mel_processing import spectrogram_torch, spec_to_mel_torch
11
+ from utils import load_wav_to_torch, load_filepaths_and_text
12
+
13
+ # import h5py
14
+
15
+
16
+ """Multi speaker version"""
17
+
18
+
19
+ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
20
+ """
21
+ 1) loads audio, speaker_id, text pairs
22
+ 2) normalizes text and converts them to sequences of integers
23
+ 3) computes spectrograms from audio files.
24
+ """
25
+
26
+ def __init__(self, audiopaths, hparams):
27
+ self.audiopaths = load_filepaths_and_text(audiopaths)
28
+ self.max_wav_value = hparams.data.max_wav_value
29
+ self.sampling_rate = hparams.data.sampling_rate
30
+ self.filter_length = hparams.data.filter_length
31
+ self.hop_length = hparams.data.hop_length
32
+ self.win_length = hparams.data.win_length
33
+ self.sampling_rate = hparams.data.sampling_rate
34
+ self.use_sr = hparams.train.use_sr
35
+ self.spec_len = hparams.train.max_speclen
36
+ self.spk_map = hparams.spk
37
+
38
+ random.seed(1234)
39
+ random.shuffle(self.audiopaths)
40
+
41
+ def get_audio(self, filename):
42
+ filename = filename.replace("\\", "/")
43
+ audio, sampling_rate = load_wav_to_torch(filename)
44
+ if sampling_rate != self.sampling_rate:
45
+ raise ValueError("{} SR doesn't match target {} SR".format(
46
+ sampling_rate, self.sampling_rate))
47
+ audio_norm = audio / self.max_wav_value
48
+ audio_norm = audio_norm.unsqueeze(0)
49
+ spec_filename = filename.replace(".wav", ".spec.pt")
50
+ if os.path.exists(spec_filename):
51
+ spec = torch.load(spec_filename)
52
+ else:
53
+ spec = spectrogram_torch(audio_norm, self.filter_length,
54
+ self.sampling_rate, self.hop_length, self.win_length,
55
+ center=False)
56
+ spec = torch.squeeze(spec, 0)
57
+ torch.save(spec, spec_filename)
58
+
59
+ spk = filename.split("/")[-2]
60
+ spk = torch.LongTensor([self.spk_map[spk]])
61
+
62
+ f0 = np.load(filename + ".f0.npy")
63
+ f0, uv = utils.interpolate_f0(f0)
64
+ f0 = torch.FloatTensor(f0)
65
+ uv = torch.FloatTensor(uv)
66
+
67
+ c = torch.load(filename+ ".soft.pt")
68
+ c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[0])
69
+
70
+
71
+ lmin = min(c.size(-1), spec.size(-1))
72
+ assert abs(c.size(-1) - spec.size(-1)) < 3, (c.size(-1), spec.size(-1), f0.shape, filename)
73
+ assert abs(audio_norm.shape[1]-lmin * self.hop_length) < 3 * self.hop_length
74
+ spec, c, f0, uv = spec[:, :lmin], c[:, :lmin], f0[:lmin], uv[:lmin]
75
+ audio_norm = audio_norm[:, :lmin * self.hop_length]
76
+ if spec.shape[1] < 60:
77
+ print("skip too short audio:", filename)
78
+ return None
79
+ if spec.shape[1] > 800:
80
+ start = random.randint(0, spec.shape[1]-800)
81
+ end = start + 790
82
+ spec, c, f0, uv = spec[:, start:end], c[:, start:end], f0[start:end], uv[start:end]
83
+ audio_norm = audio_norm[:, start * self.hop_length : end * self.hop_length]
84
+
85
+ return c, f0, spec, audio_norm, spk, uv
86
+
87
+ def __getitem__(self, index):
88
+ return self.get_audio(self.audiopaths[index][0])
89
+
90
+ def __len__(self):
91
+ return len(self.audiopaths)
92
+
93
+
94
+ class TextAudioCollate:
95
+
96
+ def __call__(self, batch):
97
+ batch = [b for b in batch if b is not None]
98
+
99
+ input_lengths, ids_sorted_decreasing = torch.sort(
100
+ torch.LongTensor([x[0].shape[1] for x in batch]),
101
+ dim=0, descending=True)
102
+
103
+ max_c_len = max([x[0].size(1) for x in batch])
104
+ max_wav_len = max([x[3].size(1) for x in batch])
105
+
106
+ lengths = torch.LongTensor(len(batch))
107
+
108
+ c_padded = torch.FloatTensor(len(batch), batch[0][0].shape[0], max_c_len)
109
+ f0_padded = torch.FloatTensor(len(batch), max_c_len)
110
+ spec_padded = torch.FloatTensor(len(batch), batch[0][2].shape[0], max_c_len)
111
+ wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len)
112
+ spkids = torch.LongTensor(len(batch), 1)
113
+ uv_padded = torch.FloatTensor(len(batch), max_c_len)
114
+
115
+ c_padded.zero_()
116
+ spec_padded.zero_()
117
+ f0_padded.zero_()
118
+ wav_padded.zero_()
119
+ uv_padded.zero_()
120
+
121
+ for i in range(len(ids_sorted_decreasing)):
122
+ row = batch[ids_sorted_decreasing[i]]
123
+
124
+ c = row[0]
125
+ c_padded[i, :, :c.size(1)] = c
126
+ lengths[i] = c.size(1)
127
+
128
+ f0 = row[1]
129
+ f0_padded[i, :f0.size(0)] = f0
130
+
131
+ spec = row[2]
132
+ spec_padded[i, :, :spec.size(1)] = spec
133
+
134
+ wav = row[3]
135
+ wav_padded[i, :, :wav.size(1)] = wav
136
+
137
+ spkids[i, 0] = row[4]
138
+
139
+ uv = row[5]
140
+ uv_padded[i, :uv.size(0)] = uv
141
+
142
+ return c_padded, f0_padded, spec_padded, wav_padded, spkids, lengths, uv_padded
flask_api.py ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import io
2
+ import logging
3
+
4
+ import soundfile
5
+ import torch
6
+ import torchaudio
7
+ from flask import Flask, request, send_file
8
+ from flask_cors import CORS
9
+
10
+ from inference.infer_tool import Svc, RealTimeVC
11
+
12
+ app = Flask(__name__)
13
+
14
+ CORS(app)
15
+
16
+ logging.getLogger('numba').setLevel(logging.WARNING)
17
+
18
+
19
+ @app.route("/voiceChangeModel", methods=["POST"])
20
+ def voice_change_model():
21
+ request_form = request.form
22
+ wave_file = request.files.get("sample", None)
23
+ # 变调信息
24
+ f_pitch_change = float(request_form.get("fPitchChange", 0))
25
+ # DAW所需的采样率
26
+ daw_sample = int(float(request_form.get("sampleRate", 0)))
27
+ speaker_id = int(float(request_form.get("sSpeakId", 0)))
28
+ # http获得wav文件并转换
29
+ input_wav_path = io.BytesIO(wave_file.read())
30
+
31
+ # 模型推理
32
+ if raw_infer:
33
+ out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path)
34
+ tar_audio = torchaudio.functional.resample(out_audio, svc_model.target_sample, daw_sample)
35
+ else:
36
+ out_audio = svc.process(svc_model, speaker_id, f_pitch_change, input_wav_path)
37
+ tar_audio = torchaudio.functional.resample(torch.from_numpy(out_audio), svc_model.target_sample, daw_sample)
38
+ # 返回音频
39
+ out_wav_path = io.BytesIO()
40
+ soundfile.write(out_wav_path, tar_audio.cpu().numpy(), daw_sample, format="wav")
41
+ out_wav_path.seek(0)
42
+ return send_file(out_wav_path, download_name="temp.wav", as_attachment=True)
43
+
44
+
45
+ if __name__ == '__main__':
46
+ # 启用则为直接切片合成,False为交叉淡化方式
47
+ # vst插件调整0.3-0.5s切片时间可以降低延迟,直接切片方法会有连接处爆音、交叉淡化会有轻微重叠声音
48
+ # 自行选择能接受的方法,或将vst最大切片时间调整为1s,此处设为Ture,延迟大音质稳定一些
49
+ raw_infer = True
50
+ # 每个模型和config是唯一对应的
51
+ model_name = "logs/32k/G_174000-Copy1.pth"
52
+ config_name = "configs/config.json"
53
+ svc_model = Svc(model_name, config_name)
54
+ svc = RealTimeVC()
55
+ # 此处与vst插件对应,不建议更改
56
+ app.run(port=6842, host="0.0.0.0", debug=False, threaded=False)
hubert/__init__.py ADDED
File without changes
hubert/__pycache__/__init__.cpython-38.pyc ADDED
Binary file (134 Bytes). View file
 
hubert/__pycache__/hubert_model.cpython-38.pyc ADDED
Binary file (7.58 kB). View file
 
hubert/checkpoint_best_legacy_500.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:60d936ec5a566776fc392e69ad8b630d14eb588111233fe313436e200a7b187b
3
+ size 1330114945
hubert/hubert_model.py ADDED
@@ -0,0 +1,222 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import random
3
+ from typing import Optional, Tuple
4
+
5
+ import torch
6
+ import torch.nn as nn
7
+ import torch.nn.functional as t_func
8
+ from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
9
+
10
+
11
+ class Hubert(nn.Module):
12
+ def __init__(self, num_label_embeddings: int = 100, mask: bool = True):
13
+ super().__init__()
14
+ self._mask = mask
15
+ self.feature_extractor = FeatureExtractor()
16
+ self.feature_projection = FeatureProjection()
17
+ self.positional_embedding = PositionalConvEmbedding()
18
+ self.norm = nn.LayerNorm(768)
19
+ self.dropout = nn.Dropout(0.1)
20
+ self.encoder = TransformerEncoder(
21
+ nn.TransformerEncoderLayer(
22
+ 768, 12, 3072, activation="gelu", batch_first=True
23
+ ),
24
+ 12,
25
+ )
26
+ self.proj = nn.Linear(768, 256)
27
+
28
+ self.masked_spec_embed = nn.Parameter(torch.FloatTensor(768).uniform_())
29
+ self.label_embedding = nn.Embedding(num_label_embeddings, 256)
30
+
31
+ def mask(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
32
+ mask = None
33
+ if self.training and self._mask:
34
+ mask = _compute_mask((x.size(0), x.size(1)), 0.8, 10, x.device, 2)
35
+ x[mask] = self.masked_spec_embed.to(x.dtype)
36
+ return x, mask
37
+
38
+ def encode(
39
+ self, x: torch.Tensor, layer: Optional[int] = None
40
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
41
+ x = self.feature_extractor(x)
42
+ x = self.feature_projection(x.transpose(1, 2))
43
+ x, mask = self.mask(x)
44
+ x = x + self.positional_embedding(x)
45
+ x = self.dropout(self.norm(x))
46
+ x = self.encoder(x, output_layer=layer)
47
+ return x, mask
48
+
49
+ def logits(self, x: torch.Tensor) -> torch.Tensor:
50
+ logits = torch.cosine_similarity(
51
+ x.unsqueeze(2),
52
+ self.label_embedding.weight.unsqueeze(0).unsqueeze(0),
53
+ dim=-1,
54
+ )
55
+ return logits / 0.1
56
+
57
+ def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
58
+ x, mask = self.encode(x)
59
+ x = self.proj(x)
60
+ logits = self.logits(x)
61
+ return logits, mask
62
+
63
+
64
+ class HubertSoft(Hubert):
65
+ def __init__(self):
66
+ super().__init__()
67
+
68
+ @torch.inference_mode()
69
+ def units(self, wav: torch.Tensor) -> torch.Tensor:
70
+ wav = t_func.pad(wav, ((400 - 320) // 2, (400 - 320) // 2))
71
+ x, _ = self.encode(wav)
72
+ return self.proj(x)
73
+
74
+
75
+ class FeatureExtractor(nn.Module):
76
+ def __init__(self):
77
+ super().__init__()
78
+ self.conv0 = nn.Conv1d(1, 512, 10, 5, bias=False)
79
+ self.norm0 = nn.GroupNorm(512, 512)
80
+ self.conv1 = nn.Conv1d(512, 512, 3, 2, bias=False)
81
+ self.conv2 = nn.Conv1d(512, 512, 3, 2, bias=False)
82
+ self.conv3 = nn.Conv1d(512, 512, 3, 2, bias=False)
83
+ self.conv4 = nn.Conv1d(512, 512, 3, 2, bias=False)
84
+ self.conv5 = nn.Conv1d(512, 512, 2, 2, bias=False)
85
+ self.conv6 = nn.Conv1d(512, 512, 2, 2, bias=False)
86
+
87
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
88
+ x = t_func.gelu(self.norm0(self.conv0(x)))
89
+ x = t_func.gelu(self.conv1(x))
90
+ x = t_func.gelu(self.conv2(x))
91
+ x = t_func.gelu(self.conv3(x))
92
+ x = t_func.gelu(self.conv4(x))
93
+ x = t_func.gelu(self.conv5(x))
94
+ x = t_func.gelu(self.conv6(x))
95
+ return x
96
+
97
+
98
+ class FeatureProjection(nn.Module):
99
+ def __init__(self):
100
+ super().__init__()
101
+ self.norm = nn.LayerNorm(512)
102
+ self.projection = nn.Linear(512, 768)
103
+ self.dropout = nn.Dropout(0.1)
104
+
105
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
106
+ x = self.norm(x)
107
+ x = self.projection(x)
108
+ x = self.dropout(x)
109
+ return x
110
+
111
+
112
+ class PositionalConvEmbedding(nn.Module):
113
+ def __init__(self):
114
+ super().__init__()
115
+ self.conv = nn.Conv1d(
116
+ 768,
117
+ 768,
118
+ kernel_size=128,
119
+ padding=128 // 2,
120
+ groups=16,
121
+ )
122
+ self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2)
123
+
124
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
125
+ x = self.conv(x.transpose(1, 2))
126
+ x = t_func.gelu(x[:, :, :-1])
127
+ return x.transpose(1, 2)
128
+
129
+
130
+ class TransformerEncoder(nn.Module):
131
+ def __init__(
132
+ self, encoder_layer: nn.TransformerEncoderLayer, num_layers: int
133
+ ) -> None:
134
+ super(TransformerEncoder, self).__init__()
135
+ self.layers = nn.ModuleList(
136
+ [copy.deepcopy(encoder_layer) for _ in range(num_layers)]
137
+ )
138
+ self.num_layers = num_layers
139
+
140
+ def forward(
141
+ self,
142
+ src: torch.Tensor,
143
+ mask: torch.Tensor = None,
144
+ src_key_padding_mask: torch.Tensor = None,
145
+ output_layer: Optional[int] = None,
146
+ ) -> torch.Tensor:
147
+ output = src
148
+ for layer in self.layers[:output_layer]:
149
+ output = layer(
150
+ output, src_mask=mask, src_key_padding_mask=src_key_padding_mask
151
+ )
152
+ return output
153
+
154
+
155
+ def _compute_mask(
156
+ shape: Tuple[int, int],
157
+ mask_prob: float,
158
+ mask_length: int,
159
+ device: torch.device,
160
+ min_masks: int = 0,
161
+ ) -> torch.Tensor:
162
+ batch_size, sequence_length = shape
163
+
164
+ if mask_length < 1:
165
+ raise ValueError("`mask_length` has to be bigger than 0.")
166
+
167
+ if mask_length > sequence_length:
168
+ raise ValueError(
169
+ f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length} and `sequence_length`: {sequence_length}`"
170
+ )
171
+
172
+ # compute number of masked spans in batch
173
+ num_masked_spans = int(mask_prob * sequence_length / mask_length + random.random())
174
+ num_masked_spans = max(num_masked_spans, min_masks)
175
+
176
+ # make sure num masked indices <= sequence_length
177
+ if num_masked_spans * mask_length > sequence_length:
178
+ num_masked_spans = sequence_length // mask_length
179
+
180
+ # SpecAugment mask to fill
181
+ mask = torch.zeros((batch_size, sequence_length), device=device, dtype=torch.bool)
182
+
183
+ # uniform distribution to sample from, make sure that offset samples are < sequence_length
184
+ uniform_dist = torch.ones(
185
+ (batch_size, sequence_length - (mask_length - 1)), device=device
186
+ )
187
+
188
+ # get random indices to mask
189
+ mask_indices = torch.multinomial(uniform_dist, num_masked_spans)
190
+
191
+ # expand masked indices to masked spans
192
+ mask_indices = (
193
+ mask_indices.unsqueeze(dim=-1)
194
+ .expand((batch_size, num_masked_spans, mask_length))
195
+ .reshape(batch_size, num_masked_spans * mask_length)
196
+ )
197
+ offsets = (
198
+ torch.arange(mask_length, device=device)[None, None, :]
199
+ .expand((batch_size, num_masked_spans, mask_length))
200
+ .reshape(batch_size, num_masked_spans * mask_length)
201
+ )
202
+ mask_idxs = mask_indices + offsets
203
+
204
+ # scatter indices to mask
205
+ mask = mask.scatter(1, mask_idxs, True)
206
+
207
+ return mask
208
+
209
+
210
+ def hubert_soft(
211
+ path: str,
212
+ ) -> HubertSoft:
213
+ r"""HuBERT-Soft from `"A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion"`.
214
+ Args:
215
+ path (str): path of a pretrained model
216
+ """
217
+ hubert = HubertSoft()
218
+ checkpoint = torch.load(path)
219
+ consume_prefix_in_state_dict_if_present(checkpoint, "module.")
220
+ hubert.load_state_dict(checkpoint)
221
+ hubert.eval()
222
+ return hubert
hubert/hubert_model_onnx.py ADDED
@@ -0,0 +1,217 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import random
3
+ from typing import Optional, Tuple
4
+
5
+ import torch
6
+ import torch.nn as nn
7
+ import torch.nn.functional as t_func
8
+ from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
9
+
10
+
11
+ class Hubert(nn.Module):
12
+ def __init__(self, num_label_embeddings: int = 100, mask: bool = True):
13
+ super().__init__()
14
+ self._mask = mask
15
+ self.feature_extractor = FeatureExtractor()
16
+ self.feature_projection = FeatureProjection()
17
+ self.positional_embedding = PositionalConvEmbedding()
18
+ self.norm = nn.LayerNorm(768)
19
+ self.dropout = nn.Dropout(0.1)
20
+ self.encoder = TransformerEncoder(
21
+ nn.TransformerEncoderLayer(
22
+ 768, 12, 3072, activation="gelu", batch_first=True
23
+ ),
24
+ 12,
25
+ )
26
+ self.proj = nn.Linear(768, 256)
27
+
28
+ self.masked_spec_embed = nn.Parameter(torch.FloatTensor(768).uniform_())
29
+ self.label_embedding = nn.Embedding(num_label_embeddings, 256)
30
+
31
+ def mask(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
32
+ mask = None
33
+ if self.training and self._mask:
34
+ mask = _compute_mask((x.size(0), x.size(1)), 0.8, 10, x.device, 2)
35
+ x[mask] = self.masked_spec_embed.to(x.dtype)
36
+ return x, mask
37
+
38
+ def encode(
39
+ self, x: torch.Tensor, layer: Optional[int] = None
40
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
41
+ x = self.feature_extractor(x)
42
+ x = self.feature_projection(x.transpose(1, 2))
43
+ x, mask = self.mask(x)
44
+ x = x + self.positional_embedding(x)
45
+ x = self.dropout(self.norm(x))
46
+ x = self.encoder(x, output_layer=layer)
47
+ return x, mask
48
+
49
+ def logits(self, x: torch.Tensor) -> torch.Tensor:
50
+ logits = torch.cosine_similarity(
51
+ x.unsqueeze(2),
52
+ self.label_embedding.weight.unsqueeze(0).unsqueeze(0),
53
+ dim=-1,
54
+ )
55
+ return logits / 0.1
56
+
57
+
58
+ class HubertSoft(Hubert):
59
+ def __init__(self):
60
+ super().__init__()
61
+
62
+ def units(self, wav: torch.Tensor) -> torch.Tensor:
63
+ wav = t_func.pad(wav, ((400 - 320) // 2, (400 - 320) // 2))
64
+ x, _ = self.encode(wav)
65
+ return self.proj(x)
66
+
67
+ def forward(self, x):
68
+ return self.units(x)
69
+
70
+ class FeatureExtractor(nn.Module):
71
+ def __init__(self):
72
+ super().__init__()
73
+ self.conv0 = nn.Conv1d(1, 512, 10, 5, bias=False)
74
+ self.norm0 = nn.GroupNorm(512, 512)
75
+ self.conv1 = nn.Conv1d(512, 512, 3, 2, bias=False)
76
+ self.conv2 = nn.Conv1d(512, 512, 3, 2, bias=False)
77
+ self.conv3 = nn.Conv1d(512, 512, 3, 2, bias=False)
78
+ self.conv4 = nn.Conv1d(512, 512, 3, 2, bias=False)
79
+ self.conv5 = nn.Conv1d(512, 512, 2, 2, bias=False)
80
+ self.conv6 = nn.Conv1d(512, 512, 2, 2, bias=False)
81
+
82
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
83
+ x = t_func.gelu(self.norm0(self.conv0(x)))
84
+ x = t_func.gelu(self.conv1(x))
85
+ x = t_func.gelu(self.conv2(x))
86
+ x = t_func.gelu(self.conv3(x))
87
+ x = t_func.gelu(self.conv4(x))
88
+ x = t_func.gelu(self.conv5(x))
89
+ x = t_func.gelu(self.conv6(x))
90
+ return x
91
+
92
+
93
+ class FeatureProjection(nn.Module):
94
+ def __init__(self):
95
+ super().__init__()
96
+ self.norm = nn.LayerNorm(512)
97
+ self.projection = nn.Linear(512, 768)
98
+ self.dropout = nn.Dropout(0.1)
99
+
100
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
101
+ x = self.norm(x)
102
+ x = self.projection(x)
103
+ x = self.dropout(x)
104
+ return x
105
+
106
+
107
+ class PositionalConvEmbedding(nn.Module):
108
+ def __init__(self):
109
+ super().__init__()
110
+ self.conv = nn.Conv1d(
111
+ 768,
112
+ 768,
113
+ kernel_size=128,
114
+ padding=128 // 2,
115
+ groups=16,
116
+ )
117
+ self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2)
118
+
119
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
120
+ x = self.conv(x.transpose(1, 2))
121
+ x = t_func.gelu(x[:, :, :-1])
122
+ return x.transpose(1, 2)
123
+
124
+
125
+ class TransformerEncoder(nn.Module):
126
+ def __init__(
127
+ self, encoder_layer: nn.TransformerEncoderLayer, num_layers: int
128
+ ) -> None:
129
+ super(TransformerEncoder, self).__init__()
130
+ self.layers = nn.ModuleList(
131
+ [copy.deepcopy(encoder_layer) for _ in range(num_layers)]
132
+ )
133
+ self.num_layers = num_layers
134
+
135
+ def forward(
136
+ self,
137
+ src: torch.Tensor,
138
+ mask: torch.Tensor = None,
139
+ src_key_padding_mask: torch.Tensor = None,
140
+ output_layer: Optional[int] = None,
141
+ ) -> torch.Tensor:
142
+ output = src
143
+ for layer in self.layers[:output_layer]:
144
+ output = layer(
145
+ output, src_mask=mask, src_key_padding_mask=src_key_padding_mask
146
+ )
147
+ return output
148
+
149
+
150
+ def _compute_mask(
151
+ shape: Tuple[int, int],
152
+ mask_prob: float,
153
+ mask_length: int,
154
+ device: torch.device,
155
+ min_masks: int = 0,
156
+ ) -> torch.Tensor:
157
+ batch_size, sequence_length = shape
158
+
159
+ if mask_length < 1:
160
+ raise ValueError("`mask_length` has to be bigger than 0.")
161
+
162
+ if mask_length > sequence_length:
163
+ raise ValueError(
164
+ f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length} and `sequence_length`: {sequence_length}`"
165
+ )
166
+
167
+ # compute number of masked spans in batch
168
+ num_masked_spans = int(mask_prob * sequence_length / mask_length + random.random())
169
+ num_masked_spans = max(num_masked_spans, min_masks)
170
+
171
+ # make sure num masked indices <= sequence_length
172
+ if num_masked_spans * mask_length > sequence_length:
173
+ num_masked_spans = sequence_length // mask_length
174
+
175
+ # SpecAugment mask to fill
176
+ mask = torch.zeros((batch_size, sequence_length), device=device, dtype=torch.bool)
177
+
178
+ # uniform distribution to sample from, make sure that offset samples are < sequence_length
179
+ uniform_dist = torch.ones(
180
+ (batch_size, sequence_length - (mask_length - 1)), device=device
181
+ )
182
+
183
+ # get random indices to mask
184
+ mask_indices = torch.multinomial(uniform_dist, num_masked_spans)
185
+
186
+ # expand masked indices to masked spans
187
+ mask_indices = (
188
+ mask_indices.unsqueeze(dim=-1)
189
+ .expand((batch_size, num_masked_spans, mask_length))
190
+ .reshape(batch_size, num_masked_spans * mask_length)
191
+ )
192
+ offsets = (
193
+ torch.arange(mask_length, device=device)[None, None, :]
194
+ .expand((batch_size, num_masked_spans, mask_length))
195
+ .reshape(batch_size, num_masked_spans * mask_length)
196
+ )
197
+ mask_idxs = mask_indices + offsets
198
+
199
+ # scatter indices to mask
200
+ mask = mask.scatter(1, mask_idxs, True)
201
+
202
+ return mask
203
+
204
+
205
+ def hubert_soft(
206
+ path: str,
207
+ ) -> HubertSoft:
208
+ r"""HuBERT-Soft from `"A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion"`.
209
+ Args:
210
+ path (str): path of a pretrained model
211
+ """
212
+ hubert = HubertSoft()
213
+ checkpoint = torch.load(path)
214
+ consume_prefix_in_state_dict_if_present(checkpoint, "module.")
215
+ hubert.load_state_dict(checkpoint)
216
+ hubert.eval()
217
+ return hubert
inference/__init__.py ADDED
File without changes
inference/__pycache__/__init__.cpython-38.pyc ADDED
Binary file (137 Bytes). View file
 
inference/__pycache__/infer_tool.cpython-38.pyc ADDED
Binary file (7.78 kB). View file
 
inference/__pycache__/slicer.cpython-38.pyc ADDED
Binary file (3.84 kB). View file
 
inference/chunks_temp.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"info": "temp_dict"}
inference/infer_tool.py ADDED
@@ -0,0 +1,233 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import hashlib
2
+ import io
3
+ import json
4
+ import logging
5
+ import os
6
+ import time
7
+ from pathlib import Path
8
+ from inference import slicer
9
+
10
+ import librosa
11
+ import numpy as np
12
+ # import onnxruntime
13
+ import parselmouth
14
+ import soundfile
15
+ import torch
16
+ import torchaudio
17
+
18
+ import cluster
19
+ from hubert import hubert_model
20
+ import utils
21
+ from models import SynthesizerTrn
22
+
23
+ logging.getLogger('matplotlib').setLevel(logging.WARNING)
24
+
25
+
26
+ def read_temp(file_name):
27
+ if not os.path.exists(file_name):
28
+ with open(file_name, "w") as f:
29
+ f.write(json.dumps({"info": "temp_dict"}))
30
+ return {}
31
+ else:
32
+ try:
33
+ with open(file_name, "r") as f:
34
+ data = f.read()
35
+ data_dict = json.loads(data)
36
+ if os.path.getsize(file_name) > 50 * 1024 * 1024:
37
+ f_name = file_name.replace("\\", "/").split("/")[-1]
38
+ print(f"clean {f_name}")
39
+ for wav_hash in list(data_dict.keys()):
40
+ if int(time.time()) - int(data_dict[wav_hash]["time"]) > 14 * 24 * 3600:
41
+ del data_dict[wav_hash]
42
+ except Exception as e:
43
+ print(e)
44
+ print(f"{file_name} error,auto rebuild file")
45
+ data_dict = {"info": "temp_dict"}
46
+ return data_dict
47
+
48
+
49
+ def write_temp(file_name, data):
50
+ with open(file_name, "w") as f:
51
+ f.write(json.dumps(data))
52
+
53
+
54
+ def timeit(func):
55
+ def run(*args, **kwargs):
56
+ t = time.time()
57
+ res = func(*args, **kwargs)
58
+ print('executing \'%s\' costed %.3fs' % (func.__name__, time.time() - t))
59
+ return res
60
+
61
+ return run
62
+
63
+
64
+ def format_wav(audio_path):
65
+ if Path(audio_path).suffix == '.wav':
66
+ return
67
+ raw_audio, raw_sample_rate = librosa.load(audio_path, mono=True, sr=None)
68
+ soundfile.write(Path(audio_path).with_suffix(".wav"), raw_audio, raw_sample_rate)
69
+
70
+
71
+ def get_end_file(dir_path, end):
72
+ file_lists = []
73
+ for root, dirs, files in os.walk(dir_path):
74
+ files = [f for f in files if f[0] != '.']
75
+ dirs[:] = [d for d in dirs if d[0] != '.']
76
+ for f_file in files:
77
+ if f_file.endswith(end):
78
+ file_lists.append(os.path.join(root, f_file).replace("\\", "/"))
79
+ return file_lists
80
+
81
+
82
+ def get_md5(content):
83
+ return hashlib.new("md5", content).hexdigest()
84
+
85
+ def fill_a_to_b(a, b):
86
+ if len(a) < len(b):
87
+ for _ in range(0, len(b) - len(a)):
88
+ a.append(a[0])
89
+
90
+ def mkdir(paths: list):
91
+ for path in paths:
92
+ if not os.path.exists(path):
93
+ os.mkdir(path)
94
+
95
+
96
+ class Svc(object):
97
+ def __init__(self, net_g_path, config_path,
98
+ device=None,
99
+ cluster_model_path="logs/44k/kmeans_10000.pt"):
100
+ self.net_g_path = net_g_path
101
+ if device is None:
102
+ self.dev = torch.device("cuda" if torch.cuda.is_available() else "cpu")
103
+ else:
104
+ self.dev = torch.device(device)
105
+ self.net_g_ms = None
106
+ self.hps_ms = utils.get_hparams_from_file(config_path)
107
+ self.target_sample = self.hps_ms.data.sampling_rate
108
+ self.hop_size = self.hps_ms.data.hop_length
109
+ self.spk2id = self.hps_ms.spk
110
+ # 加载hubert
111
+ self.hubert_model = utils.get_hubert_model().to(self.dev)
112
+ self.load_model()
113
+ if os.path.exists(cluster_model_path):
114
+ self.cluster_model = cluster.get_cluster_model(cluster_model_path)
115
+
116
+ def load_model(self):
117
+ # 获取模型配置
118
+ self.net_g_ms = SynthesizerTrn(
119
+ self.hps_ms.data.filter_length // 2 + 1,
120
+ self.hps_ms.train.segment_size // self.hps_ms.data.hop_length,
121
+ **self.hps_ms.model)
122
+ _ = utils.load_checkpoint(self.net_g_path, self.net_g_ms, None)
123
+ if "half" in self.net_g_path and torch.cuda.is_available():
124
+ _ = self.net_g_ms.half().eval().to(self.dev)
125
+ else:
126
+ _ = self.net_g_ms.eval().to(self.dev)
127
+
128
+
129
+
130
+ def get_unit_f0(self, in_path, tran, cluster_infer_ratio, speaker):
131
+
132
+ wav, sr = librosa.load(in_path, sr=self.target_sample)
133
+
134
+ f0 = utils.compute_f0_parselmouth(wav, sampling_rate=self.target_sample, hop_length=self.hop_size)
135
+ f0, uv = utils.interpolate_f0(f0)
136
+ f0 = torch.FloatTensor(f0)
137
+ uv = torch.FloatTensor(uv)
138
+ f0 = f0 * 2 ** (tran / 12)
139
+ f0 = f0.unsqueeze(0).to(self.dev)
140
+ uv = uv.unsqueeze(0).to(self.dev)
141
+
142
+ wav16k = librosa.resample(wav, orig_sr=self.target_sample, target_sr=16000)
143
+ wav16k = torch.from_numpy(wav16k).to(self.dev)
144
+ c = utils.get_hubert_content(self.hubert_model, wav_16k_tensor=wav16k)
145
+ c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[1])
146
+
147
+ if cluster_infer_ratio !=0:
148
+ cluster_c = cluster.get_cluster_center_result(self.cluster_model, c.numpy().T, speaker).T
149
+ cluster_c = torch.FloatTensor(cluster_c)
150
+ c = cluster_infer_ratio * cluster_c + (1 - cluster_infer_ratio) * c
151
+
152
+ c = c.unsqueeze(0)
153
+ return c, f0, uv
154
+
155
+ def infer(self, speaker, tran, raw_path,
156
+ cluster_infer_ratio=0,
157
+ auto_predict_f0=False,
158
+ noice_scale=0.4):
159
+ speaker_id = self.spk2id[speaker]
160
+ sid = torch.LongTensor([int(speaker_id)]).to(self.dev).unsqueeze(0)
161
+ c, f0, uv = self.get_unit_f0(raw_path, tran, cluster_infer_ratio, speaker)
162
+ if "half" in self.net_g_path and torch.cuda.is_available():
163
+ c = c.half()
164
+ with torch.no_grad():
165
+ start = time.time()
166
+ audio = self.net_g_ms.infer(c, f0=f0, g=sid, uv=uv, predict_f0=auto_predict_f0, noice_scale=noice_scale)[0,0].data.float()
167
+ use_time = time.time() - start
168
+ print("vits use time:{}".format(use_time))
169
+ return audio, audio.shape[-1]
170
+
171
+ def slice_inference(self,raw_audio_path, spk, tran, slice_db,cluster_infer_ratio, auto_predict_f0,noice_scale, pad_seconds=0.5):
172
+ wav_path = raw_audio_path
173
+ chunks = slicer.cut(wav_path, db_thresh=slice_db)
174
+ audio_data, audio_sr = slicer.chunks2audio(wav_path, chunks)
175
+
176
+ audio = []
177
+ for (slice_tag, data) in audio_data:
178
+ print(f'#=====segment start, {round(len(data) / audio_sr, 3)}s======')
179
+ # padd
180
+ pad_len = int(audio_sr * pad_seconds)
181
+ data = np.concatenate([np.zeros([pad_len]), data, np.zeros([pad_len])])
182
+ length = int(np.ceil(len(data) / audio_sr * self.target_sample))
183
+ raw_path = io.BytesIO()
184
+ soundfile.write(raw_path, data, audio_sr, format="wav")
185
+ raw_path.seek(0)
186
+ if slice_tag:
187
+ print('jump empty segment')
188
+ _audio = np.zeros(length)
189
+ else:
190
+ out_audio, out_sr = self.infer(spk, tran, raw_path,
191
+ cluster_infer_ratio=cluster_infer_ratio,
192
+ auto_predict_f0=auto_predict_f0,
193
+ noice_scale=noice_scale
194
+ )
195
+ _audio = out_audio.cpu().numpy()
196
+
197
+ pad_len = int(self.target_sample * pad_seconds)
198
+ _audio = _audio[pad_len:-pad_len]
199
+ audio.extend(list(_audio))
200
+ return np.array(audio)
201
+
202
+
203
+ class RealTimeVC:
204
+ def __init__(self):
205
+ self.last_chunk = None
206
+ self.last_o = None
207
+ self.chunk_len = 16000 # 区块长度
208
+ self.pre_len = 3840 # 交叉淡化长度,640的倍数
209
+
210
+ """输入输出都是1维numpy 音频波形数组"""
211
+
212
+ def process(self, svc_model, speaker_id, f_pitch_change, input_wav_path):
213
+ import maad
214
+ audio, sr = torchaudio.load(input_wav_path)
215
+ audio = audio.cpu().numpy()[0]
216
+ temp_wav = io.BytesIO()
217
+ if self.last_chunk is None:
218
+ input_wav_path.seek(0)
219
+ audio, sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path)
220
+ audio = audio.cpu().numpy()
221
+ self.last_chunk = audio[-self.pre_len:]
222
+ self.last_o = audio
223
+ return audio[-self.chunk_len:]
224
+ else:
225
+ audio = np.concatenate([self.last_chunk, audio])
226
+ soundfile.write(temp_wav, audio, sr, format="wav")
227
+ temp_wav.seek(0)
228
+ audio, sr = svc_model.infer(speaker_id, f_pitch_change, temp_wav)
229
+ audio = audio.cpu().numpy()
230
+ ret = maad.util.crossfade(self.last_o, audio, self.pre_len)
231
+ self.last_chunk = audio[-self.pre_len:]
232
+ self.last_o = audio
233
+ return ret[self.chunk_len:2 * self.chunk_len]
inference/infer_tool_grad.py ADDED
@@ -0,0 +1,160 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import hashlib
2
+ import json
3
+ import logging
4
+ import os
5
+ import time
6
+ from pathlib import Path
7
+ import io
8
+ import librosa
9
+ import maad
10
+ import numpy as np
11
+ from inference import slicer
12
+ import parselmouth
13
+ import soundfile
14
+ import torch
15
+ import torchaudio
16
+
17
+ from hubert import hubert_model
18
+ import utils
19
+ from models import SynthesizerTrn
20
+ logging.getLogger('numba').setLevel(logging.WARNING)
21
+ logging.getLogger('matplotlib').setLevel(logging.WARNING)
22
+
23
+ def resize2d_f0(x, target_len):
24
+ source = np.array(x)
25
+ source[source < 0.001] = np.nan
26
+ target = np.interp(np.arange(0, len(source) * target_len, len(source)) / target_len, np.arange(0, len(source)),
27
+ source)
28
+ res = np.nan_to_num(target)
29
+ return res
30
+
31
+ def get_f0(x, p_len,f0_up_key=0):
32
+
33
+ time_step = 160 / 16000 * 1000
34
+ f0_min = 50
35
+ f0_max = 1100
36
+ f0_mel_min = 1127 * np.log(1 + f0_min / 700)
37
+ f0_mel_max = 1127 * np.log(1 + f0_max / 700)
38
+
39
+ f0 = parselmouth.Sound(x, 16000).to_pitch_ac(
40
+ time_step=time_step / 1000, voicing_threshold=0.6,
41
+ pitch_floor=f0_min, pitch_ceiling=f0_max).selected_array['frequency']
42
+
43
+ pad_size=(p_len - len(f0) + 1) // 2
44
+ if(pad_size>0 or p_len - len(f0) - pad_size>0):
45
+ f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
46
+
47
+ f0 *= pow(2, f0_up_key / 12)
48
+ f0_mel = 1127 * np.log(1 + f0 / 700)
49
+ f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (f0_mel_max - f0_mel_min) + 1
50
+ f0_mel[f0_mel <= 1] = 1
51
+ f0_mel[f0_mel > 255] = 255
52
+ f0_coarse = np.rint(f0_mel).astype(np.int)
53
+ return f0_coarse, f0
54
+
55
+ def clean_pitch(input_pitch):
56
+ num_nan = np.sum(input_pitch == 1)
57
+ if num_nan / len(input_pitch) > 0.9:
58
+ input_pitch[input_pitch != 1] = 1
59
+ return input_pitch
60
+
61
+
62
+ def plt_pitch(input_pitch):
63
+ input_pitch = input_pitch.astype(float)
64
+ input_pitch[input_pitch == 1] = np.nan
65
+ return input_pitch
66
+
67
+
68
+ def f0_to_pitch(ff):
69
+ f0_pitch = 69 + 12 * np.log2(ff / 440)
70
+ return f0_pitch
71
+
72
+
73
+ def fill_a_to_b(a, b):
74
+ if len(a) < len(b):
75
+ for _ in range(0, len(b) - len(a)):
76
+ a.append(a[0])
77
+
78
+
79
+ def mkdir(paths: list):
80
+ for path in paths:
81
+ if not os.path.exists(path):
82
+ os.mkdir(path)
83
+
84
+
85
+ class VitsSvc(object):
86
+ def __init__(self):
87
+ self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
88
+ self.SVCVITS = None
89
+ self.hps = None
90
+ self.speakers = None
91
+ self.hubert_soft = utils.get_hubert_model()
92
+
93
+ def set_device(self, device):
94
+ self.device = torch.device(device)
95
+ self.hubert_soft.to(self.device)
96
+ if self.SVCVITS != None:
97
+ self.SVCVITS.to(self.device)
98
+
99
+ def loadCheckpoint(self, path):
100
+ self.hps = utils.get_hparams_from_file(f"checkpoints/{path}/config.json")
101
+ self.SVCVITS = SynthesizerTrn(
102
+ self.hps.data.filter_length // 2 + 1,
103
+ self.hps.train.segment_size // self.hps.data.hop_length,
104
+ **self.hps.model)
105
+ _ = utils.load_checkpoint(f"checkpoints/{path}/model.pth", self.SVCVITS, None)
106
+ _ = self.SVCVITS.eval().to(self.device)
107
+ self.speakers = self.hps.spk
108
+
109
+ def get_units(self, source, sr):
110
+ source = source.unsqueeze(0).to(self.device)
111
+ with torch.inference_mode():
112
+ units = self.hubert_soft.units(source)
113
+ return units
114
+
115
+
116
+ def get_unit_pitch(self, in_path, tran):
117
+ source, sr = torchaudio.load(in_path)
118
+ source = torchaudio.functional.resample(source, sr, 16000)
119
+ if len(source.shape) == 2 and source.shape[1] >= 2:
120
+ source = torch.mean(source, dim=0).unsqueeze(0)
121
+ soft = self.get_units(source, sr).squeeze(0).cpu().numpy()
122
+ f0_coarse, f0 = get_f0(source.cpu().numpy()[0], soft.shape[0]*2, tran)
123
+ return soft, f0
124
+
125
+ def infer(self, speaker_id, tran, raw_path):
126
+ speaker_id = self.speakers[speaker_id]
127
+ sid = torch.LongTensor([int(speaker_id)]).to(self.device).unsqueeze(0)
128
+ soft, pitch = self.get_unit_pitch(raw_path, tran)
129
+ f0 = torch.FloatTensor(clean_pitch(pitch)).unsqueeze(0).to(self.device)
130
+ stn_tst = torch.FloatTensor(soft)
131
+ with torch.no_grad():
132
+ x_tst = stn_tst.unsqueeze(0).to(self.device)
133
+ x_tst = torch.repeat_interleave(x_tst, repeats=2, dim=1).transpose(1, 2)
134
+ audio = self.SVCVITS.infer(x_tst, f0=f0, g=sid)[0,0].data.float()
135
+ return audio, audio.shape[-1]
136
+
137
+ def inference(self,srcaudio,chara,tran,slice_db):
138
+ sampling_rate, audio = srcaudio
139
+ audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
140
+ if len(audio.shape) > 1:
141
+ audio = librosa.to_mono(audio.transpose(1, 0))
142
+ if sampling_rate != 16000:
143
+ audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
144
+ soundfile.write("tmpwav.wav", audio, 16000, format="wav")
145
+ chunks = slicer.cut("tmpwav.wav", db_thresh=slice_db)
146
+ audio_data, audio_sr = slicer.chunks2audio("tmpwav.wav", chunks)
147
+ audio = []
148
+ for (slice_tag, data) in audio_data:
149
+ length = int(np.ceil(len(data) / audio_sr * self.hps.data.sampling_rate))
150
+ raw_path = io.BytesIO()
151
+ soundfile.write(raw_path, data, audio_sr, format="wav")
152
+ raw_path.seek(0)
153
+ if slice_tag:
154
+ _audio = np.zeros(length)
155
+ else:
156
+ out_audio, out_sr = self.infer(chara, tran, raw_path)
157
+ _audio = out_audio.cpu().numpy()
158
+ audio.extend(list(_audio))
159
+ audio = (np.array(audio) * 32768.0).astype('int16')
160
+ return (self.hps.data.sampling_rate,audio)
inference/slicer.py ADDED
@@ -0,0 +1,142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import librosa
2
+ import torch
3
+ import torchaudio
4
+
5
+
6
+ class Slicer:
7
+ def __init__(self,
8
+ sr: int,
9
+ threshold: float = -40.,
10
+ min_length: int = 5000,
11
+ min_interval: int = 300,
12
+ hop_size: int = 20,
13
+ max_sil_kept: int = 5000):
14
+ if not min_length >= min_interval >= hop_size:
15
+ raise ValueError('The following condition must be satisfied: min_length >= min_interval >= hop_size')
16
+ if not max_sil_kept >= hop_size:
17
+ raise ValueError('The following condition must be satisfied: max_sil_kept >= hop_size')
18
+ min_interval = sr * min_interval / 1000
19
+ self.threshold = 10 ** (threshold / 20.)
20
+ self.hop_size = round(sr * hop_size / 1000)
21
+ self.win_size = min(round(min_interval), 4 * self.hop_size)
22
+ self.min_length = round(sr * min_length / 1000 / self.hop_size)
23
+ self.min_interval = round(min_interval / self.hop_size)
24
+ self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size)
25
+
26
+ def _apply_slice(self, waveform, begin, end):
27
+ if len(waveform.shape) > 1:
28
+ return waveform[:, begin * self.hop_size: min(waveform.shape[1], end * self.hop_size)]
29
+ else:
30
+ return waveform[begin * self.hop_size: min(waveform.shape[0], end * self.hop_size)]
31
+
32
+ # @timeit
33
+ def slice(self, waveform):
34
+ if len(waveform.shape) > 1:
35
+ samples = librosa.to_mono(waveform)
36
+ else:
37
+ samples = waveform
38
+ if samples.shape[0] <= self.min_length:
39
+ return {"0": {"slice": False, "split_time": f"0,{len(waveform)}"}}
40
+ rms_list = librosa.feature.rms(y=samples, frame_length=self.win_size, hop_length=self.hop_size).squeeze(0)
41
+ sil_tags = []
42
+ silence_start = None
43
+ clip_start = 0
44
+ for i, rms in enumerate(rms_list):
45
+ # Keep looping while frame is silent.
46
+ if rms < self.threshold:
47
+ # Record start of silent frames.
48
+ if silence_start is None:
49
+ silence_start = i
50
+ continue
51
+ # Keep looping while frame is not silent and silence start has not been recorded.
52
+ if silence_start is None:
53
+ continue
54
+ # Clear recorded silence start if interval is not enough or clip is too short
55
+ is_leading_silence = silence_start == 0 and i > self.max_sil_kept
56
+ need_slice_middle = i - silence_start >= self.min_interval and i - clip_start >= self.min_length
57
+ if not is_leading_silence and not need_slice_middle:
58
+ silence_start = None
59
+ continue
60
+ # Need slicing. Record the range of silent frames to be removed.
61
+ if i - silence_start <= self.max_sil_kept:
62
+ pos = rms_list[silence_start: i + 1].argmin() + silence_start
63
+ if silence_start == 0:
64
+ sil_tags.append((0, pos))
65
+ else:
66
+ sil_tags.append((pos, pos))
67
+ clip_start = pos
68
+ elif i - silence_start <= self.max_sil_kept * 2:
69
+ pos = rms_list[i - self.max_sil_kept: silence_start + self.max_sil_kept + 1].argmin()
70
+ pos += i - self.max_sil_kept
71
+ pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start
72
+ pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept
73
+ if silence_start == 0:
74
+ sil_tags.append((0, pos_r))
75
+ clip_start = pos_r
76
+ else:
77
+ sil_tags.append((min(pos_l, pos), max(pos_r, pos)))
78
+ clip_start = max(pos_r, pos)
79
+ else:
80
+ pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start
81
+ pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept
82
+ if silence_start == 0:
83
+ sil_tags.append((0, pos_r))
84
+ else:
85
+ sil_tags.append((pos_l, pos_r))
86
+ clip_start = pos_r
87
+ silence_start = None
88
+ # Deal with trailing silence.
89
+ total_frames = rms_list.shape[0]
90
+ if silence_start is not None and total_frames - silence_start >= self.min_interval:
91
+ silence_end = min(total_frames, silence_start + self.max_sil_kept)
92
+ pos = rms_list[silence_start: silence_end + 1].argmin() + silence_start
93
+ sil_tags.append((pos, total_frames + 1))
94
+ # Apply and return slices.
95
+ if len(sil_tags) == 0:
96
+ return {"0": {"slice": False, "split_time": f"0,{len(waveform)}"}}
97
+ else:
98
+ chunks = []
99
+ # 第一段静音并非从头开始,补上有声片段
100
+ if sil_tags[0][0]:
101
+ chunks.append(
102
+ {"slice": False, "split_time": f"0,{min(waveform.shape[0], sil_tags[0][0] * self.hop_size)}"})
103
+ for i in range(0, len(sil_tags)):
104
+ # 标识有声片段(跳过第一段)
105
+ if i:
106
+ chunks.append({"slice": False,
107
+ "split_time": f"{sil_tags[i - 1][1] * self.hop_size},{min(waveform.shape[0], sil_tags[i][0] * self.hop_size)}"})
108
+ # 标识所有静音片段
109
+ chunks.append({"slice": True,
110
+ "split_time": f"{sil_tags[i][0] * self.hop_size},{min(waveform.shape[0], sil_tags[i][1] * self.hop_size)}"})
111
+ # 最后一段静音并非结尾,补上结尾片段
112
+ if sil_tags[-1][1] * self.hop_size < len(waveform):
113
+ chunks.append({"slice": False, "split_time": f"{sil_tags[-1][1] * self.hop_size},{len(waveform)}"})
114
+ chunk_dict = {}
115
+ for i in range(len(chunks)):
116
+ chunk_dict[str(i)] = chunks[i]
117
+ return chunk_dict
118
+
119
+
120
+ def cut(audio_path, db_thresh=-30, min_len=5000):
121
+ audio, sr = librosa.load(audio_path, sr=None)
122
+ slicer = Slicer(
123
+ sr=sr,
124
+ threshold=db_thresh,
125
+ min_length=min_len
126
+ )
127
+ chunks = slicer.slice(audio)
128
+ return chunks
129
+
130
+
131
+ def chunks2audio(audio_path, chunks):
132
+ chunks = dict(chunks)
133
+ audio, sr = torchaudio.load(audio_path)
134
+ if len(audio.shape) == 2 and audio.shape[1] >= 2:
135
+ audio = torch.mean(audio, dim=0).unsqueeze(0)
136
+ audio = audio.cpu().numpy()[0]
137
+ result = []
138
+ for k, v in chunks.items():
139
+ tag = v["split_time"].split(",")
140
+ if tag[0] != tag[1]:
141
+ result.append((v["slice"], audio[int(tag[0]):int(tag[1])]))
142
+ return result, sr
inference_main.py ADDED
@@ -0,0 +1,100 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import io
2
+ import logging
3
+ import time
4
+ from pathlib import Path
5
+
6
+ import librosa
7
+ import matplotlib.pyplot as plt
8
+ import numpy as np
9
+ import soundfile
10
+
11
+ from inference import infer_tool
12
+ from inference import slicer
13
+ from inference.infer_tool import Svc
14
+
15
+ logging.getLogger('numba').setLevel(logging.WARNING)
16
+ chunks_dict = infer_tool.read_temp("inference/chunks_temp.json")
17
+
18
+
19
+
20
+ def main():
21
+ import argparse
22
+
23
+ parser = argparse.ArgumentParser(description='sovits4 inference')
24
+
25
+ # 一定要设置的部分
26
+ parser.add_argument('-m', '--model_path', type=str, default="/Volumes/Extend/下载/G_20800.pth", help='模型路径')
27
+ parser.add_argument('-c', '--config_path', type=str, default="configs/config.json", help='配置文件路径')
28
+ parser.add_argument('-n', '--clean_names', type=str, nargs='+', default=["君の知らない物語-src"], help='wav文件名列表,放在raw文件夹下')
29
+ parser.add_argument('-t', '--trans', type=int, nargs='+', default=[0], help='音高调整,支持正负(半音)')
30
+ parser.add_argument('-s', '--spk_list', type=str, nargs='+', default=['nyaru'], help='合成目标说话人名称')
31
+
32
+ # 可选项部分
33
+ parser.add_argument('-a', '--auto_predict_f0', action='store_true', default=False,
34
+ help='语音转换自动预测音高,转换歌声时不要打开这个会严重跑调')
35
+ parser.add_argument('-cm', '--cluster_model_path', type=str, default="/Volumes/Extend/下载/so-vits-svc-4.0/logs/44k/kmeans_10000.pt", help='聚类模型路径,如果没有训练聚类则随便填')
36
+ parser.add_argument('-cr', '--cluster_infer_ratio', type=float, default=1, help='聚类方案占比,范围0-1,若没有训练聚类模型则填0即可')
37
+
38
+ # 不用动的部分
39
+ parser.add_argument('-sd', '--slice_db', type=int, default=-40, help='默认-40,嘈杂的音频可以-30,干声保留呼吸可以-50')
40
+ parser.add_argument('-d', '--device', type=str, default=None, help='推理设备,None则为自动选择cpu和gpu')
41
+ parser.add_argument('-ns', '--noice_scale', type=float, default=0.4, help='噪音级别,会影响咬字和音质,较为玄学')
42
+ parser.add_argument('-p', '--pad_seconds', type=float, default=0.5, help='推理音频pad秒数,由于未知原因开头结尾会有异响,pad一小段静音段后就不会出现')
43
+ parser.add_argument('-wf', '--wav_format', type=str, default='flac', help='音频输出格式')
44
+
45
+ args = parser.parse_args()
46
+
47
+ svc_model = Svc(args.model_path, args.config_path, args.device, args.cluster_model_path)
48
+ infer_tool.mkdir(["raw", "results"])
49
+ clean_names = args.clean_names
50
+ trans = args.trans
51
+ spk_list = args.spk_list
52
+ slice_db = args.slice_db
53
+ wav_format = args.wav_format
54
+ auto_predict_f0 = args.auto_predict_f0
55
+ cluster_infer_ratio = args.cluster_infer_ratio
56
+ noice_scale = args.noice_scale
57
+ pad_seconds = args.pad_seconds
58
+
59
+ infer_tool.fill_a_to_b(trans, clean_names)
60
+ for clean_name, tran in zip(clean_names, trans):
61
+ raw_audio_path = f"raw/{clean_name}"
62
+ if "." not in raw_audio_path:
63
+ raw_audio_path += ".wav"
64
+ infer_tool.format_wav(raw_audio_path)
65
+ wav_path = Path(raw_audio_path).with_suffix('.wav')
66
+ chunks = slicer.cut(wav_path, db_thresh=slice_db)
67
+ audio_data, audio_sr = slicer.chunks2audio(wav_path, chunks)
68
+
69
+ for spk in spk_list:
70
+ audio = []
71
+ for (slice_tag, data) in audio_data:
72
+ print(f'#=====segment start, {round(len(data) / audio_sr, 3)}s======')
73
+ # padd
74
+ pad_len = int(audio_sr * pad_seconds)
75
+ data = np.concatenate([np.zeros([pad_len]), data, np.zeros([pad_len])])
76
+ length = int(np.ceil(len(data) / audio_sr * svc_model.target_sample))
77
+ raw_path = io.BytesIO()
78
+ soundfile.write(raw_path, data, audio_sr, format="wav")
79
+ raw_path.seek(0)
80
+ if slice_tag:
81
+ print('jump empty segment')
82
+ _audio = np.zeros(length)
83
+ else:
84
+ out_audio, out_sr = svc_model.infer(spk, tran, raw_path,
85
+ cluster_infer_ratio=cluster_infer_ratio,
86
+ auto_predict_f0=auto_predict_f0,
87
+ noice_scale=noice_scale
88
+ )
89
+ _audio = out_audio.cpu().numpy()
90
+
91
+ pad_len = int(svc_model.target_sample * pad_seconds)
92
+ _audio = _audio[pad_len:-pad_len]
93
+ audio.extend(list(_audio))
94
+ key = "auto" if auto_predict_f0 else f"{tran}key"
95
+ cluster_name = "" if cluster_infer_ratio == 0 else f"_{cluster_infer_ratio}"
96
+ res_path = f'./results/old——{clean_name}_{key}_{spk}{cluster_name}.{wav_format}'
97
+ soundfile.write(res_path, audio, svc_model.target_sample, format=wav_format)
98
+
99
+ if __name__ == '__main__':
100
+ main()
models.py ADDED
@@ -0,0 +1,420 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import math
3
+ import torch
4
+ from torch import nn
5
+ from torch.nn import functional as F
6
+
7
+ import modules.attentions as attentions
8
+ import modules.commons as commons
9
+ import modules.modules as modules
10
+
11
+ from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
12
+ from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
13
+
14
+ import utils
15
+ from modules.commons import init_weights, get_padding
16
+ from vdecoder.hifigan.models import Generator
17
+ from utils import f0_to_coarse
18
+
19
+ class ResidualCouplingBlock(nn.Module):
20
+ def __init__(self,
21
+ channels,
22
+ hidden_channels,
23
+ kernel_size,
24
+ dilation_rate,
25
+ n_layers,
26
+ n_flows=4,
27
+ gin_channels=0):
28
+ super().__init__()
29
+ self.channels = channels
30
+ self.hidden_channels = hidden_channels
31
+ self.kernel_size = kernel_size
32
+ self.dilation_rate = dilation_rate
33
+ self.n_layers = n_layers
34
+ self.n_flows = n_flows
35
+ self.gin_channels = gin_channels
36
+
37
+ self.flows = nn.ModuleList()
38
+ for i in range(n_flows):
39
+ self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
40
+ self.flows.append(modules.Flip())
41
+
42
+ def forward(self, x, x_mask, g=None, reverse=False):
43
+ if not reverse:
44
+ for flow in self.flows:
45
+ x, _ = flow(x, x_mask, g=g, reverse=reverse)
46
+ else:
47
+ for flow in reversed(self.flows):
48
+ x = flow(x, x_mask, g=g, reverse=reverse)
49
+ return x
50
+
51
+
52
+ class Encoder(nn.Module):
53
+ def __init__(self,
54
+ in_channels,
55
+ out_channels,
56
+ hidden_channels,
57
+ kernel_size,
58
+ dilation_rate,
59
+ n_layers,
60
+ gin_channels=0):
61
+ super().__init__()
62
+ self.in_channels = in_channels
63
+ self.out_channels = out_channels
64
+ self.hidden_channels = hidden_channels
65
+ self.kernel_size = kernel_size
66
+ self.dilation_rate = dilation_rate
67
+ self.n_layers = n_layers
68
+ self.gin_channels = gin_channels
69
+
70
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
71
+ self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
72
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
73
+
74
+ def forward(self, x, x_lengths, g=None):
75
+ # print(x.shape,x_lengths.shape)
76
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
77
+ x = self.pre(x) * x_mask
78
+ x = self.enc(x, x_mask, g=g)
79
+ stats = self.proj(x) * x_mask
80
+ m, logs = torch.split(stats, self.out_channels, dim=1)
81
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
82
+ return z, m, logs, x_mask
83
+
84
+
85
+ class TextEncoder(nn.Module):
86
+ def __init__(self,
87
+ out_channels,
88
+ hidden_channels,
89
+ kernel_size,
90
+ n_layers,
91
+ gin_channels=0,
92
+ filter_channels=None,
93
+ n_heads=None,
94
+ p_dropout=None):
95
+ super().__init__()
96
+ self.out_channels = out_channels
97
+ self.hidden_channels = hidden_channels
98
+ self.kernel_size = kernel_size
99
+ self.n_layers = n_layers
100
+ self.gin_channels = gin_channels
101
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
102
+ self.f0_emb = nn.Embedding(256, hidden_channels)
103
+
104
+ self.enc_ = attentions.Encoder(
105
+ hidden_channels,
106
+ filter_channels,
107
+ n_heads,
108
+ n_layers,
109
+ kernel_size,
110
+ p_dropout)
111
+
112
+ def forward(self, x, x_mask, f0=None, noice_scale=1):
113
+ x = x + self.f0_emb(f0).transpose(1,2)
114
+ x = self.enc_(x * x_mask, x_mask)
115
+ stats = self.proj(x) * x_mask
116
+ m, logs = torch.split(stats, self.out_channels, dim=1)
117
+ z = (m + torch.randn_like(m) * torch.exp(logs) * noice_scale) * x_mask
118
+
119
+ return z, m, logs, x_mask
120
+
121
+
122
+
123
+ class DiscriminatorP(torch.nn.Module):
124
+ def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
125
+ super(DiscriminatorP, self).__init__()
126
+ self.period = period
127
+ self.use_spectral_norm = use_spectral_norm
128
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
129
+ self.convs = nn.ModuleList([
130
+ norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
131
+ norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
132
+ norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
133
+ norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
134
+ norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
135
+ ])
136
+ self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
137
+
138
+ def forward(self, x):
139
+ fmap = []
140
+
141
+ # 1d to 2d
142
+ b, c, t = x.shape
143
+ if t % self.period != 0: # pad first
144
+ n_pad = self.period - (t % self.period)
145
+ x = F.pad(x, (0, n_pad), "reflect")
146
+ t = t + n_pad
147
+ x = x.view(b, c, t // self.period, self.period)
148
+
149
+ for l in self.convs:
150
+ x = l(x)
151
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
152
+ fmap.append(x)
153
+ x = self.conv_post(x)
154
+ fmap.append(x)
155
+ x = torch.flatten(x, 1, -1)
156
+
157
+ return x, fmap
158
+
159
+
160
+ class DiscriminatorS(torch.nn.Module):
161
+ def __init__(self, use_spectral_norm=False):
162
+ super(DiscriminatorS, self).__init__()
163
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
164
+ self.convs = nn.ModuleList([
165
+ norm_f(Conv1d(1, 16, 15, 1, padding=7)),
166
+ norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
167
+ norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
168
+ norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
169
+ norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
170
+ norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
171
+ ])
172
+ self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
173
+
174
+ def forward(self, x):
175
+ fmap = []
176
+
177
+ for l in self.convs:
178
+ x = l(x)
179
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
180
+ fmap.append(x)
181
+ x = self.conv_post(x)
182
+ fmap.append(x)
183
+ x = torch.flatten(x, 1, -1)
184
+
185
+ return x, fmap
186
+
187
+
188
+ class MultiPeriodDiscriminator(torch.nn.Module):
189
+ def __init__(self, use_spectral_norm=False):
190
+ super(MultiPeriodDiscriminator, self).__init__()
191
+ periods = [2,3,5,7,11]
192
+
193
+ discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
194
+ discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
195
+ self.discriminators = nn.ModuleList(discs)
196
+
197
+ def forward(self, y, y_hat):
198
+ y_d_rs = []
199
+ y_d_gs = []
200
+ fmap_rs = []
201
+ fmap_gs = []
202
+ for i, d in enumerate(self.discriminators):
203
+ y_d_r, fmap_r = d(y)
204
+ y_d_g, fmap_g = d(y_hat)
205
+ y_d_rs.append(y_d_r)
206
+ y_d_gs.append(y_d_g)
207
+ fmap_rs.append(fmap_r)
208
+ fmap_gs.append(fmap_g)
209
+
210
+ return y_d_rs, y_d_gs, fmap_rs, fmap_gs
211
+
212
+
213
+ class SpeakerEncoder(torch.nn.Module):
214
+ def __init__(self, mel_n_channels=80, model_num_layers=3, model_hidden_size=256, model_embedding_size=256):
215
+ super(SpeakerEncoder, self).__init__()
216
+ self.lstm = nn.LSTM(mel_n_channels, model_hidden_size, model_num_layers, batch_first=True)
217
+ self.linear = nn.Linear(model_hidden_size, model_embedding_size)
218
+ self.relu = nn.ReLU()
219
+
220
+ def forward(self, mels):
221
+ self.lstm.flatten_parameters()
222
+ _, (hidden, _) = self.lstm(mels)
223
+ embeds_raw = self.relu(self.linear(hidden[-1]))
224
+ return embeds_raw / torch.norm(embeds_raw, dim=1, keepdim=True)
225
+
226
+ def compute_partial_slices(self, total_frames, partial_frames, partial_hop):
227
+ mel_slices = []
228
+ for i in range(0, total_frames-partial_frames, partial_hop):
229
+ mel_range = torch.arange(i, i+partial_frames)
230
+ mel_slices.append(mel_range)
231
+
232
+ return mel_slices
233
+
234
+ def embed_utterance(self, mel, partial_frames=128, partial_hop=64):
235
+ mel_len = mel.size(1)
236
+ last_mel = mel[:,-partial_frames:]
237
+
238
+ if mel_len > partial_frames:
239
+ mel_slices = self.compute_partial_slices(mel_len, partial_frames, partial_hop)
240
+ mels = list(mel[:,s] for s in mel_slices)
241
+ mels.append(last_mel)
242
+ mels = torch.stack(tuple(mels), 0).squeeze(1)
243
+
244
+ with torch.no_grad():
245
+ partial_embeds = self(mels)
246
+ embed = torch.mean(partial_embeds, axis=0).unsqueeze(0)
247
+ #embed = embed / torch.linalg.norm(embed, 2)
248
+ else:
249
+ with torch.no_grad():
250
+ embed = self(last_mel)
251
+
252
+ return embed
253
+
254
+ class F0Decoder(nn.Module):
255
+ def __init__(self,
256
+ out_channels,
257
+ hidden_channels,
258
+ filter_channels,
259
+ n_heads,
260
+ n_layers,
261
+ kernel_size,
262
+ p_dropout,
263
+ spk_channels=0):
264
+ super().__init__()
265
+ self.out_channels = out_channels
266
+ self.hidden_channels = hidden_channels
267
+ self.filter_channels = filter_channels
268
+ self.n_heads = n_heads
269
+ self.n_layers = n_layers
270
+ self.kernel_size = kernel_size
271
+ self.p_dropout = p_dropout
272
+ self.spk_channels = spk_channels
273
+
274
+ self.prenet = nn.Conv1d(hidden_channels, hidden_channels, 3, padding=1)
275
+ self.decoder = attentions.FFT(
276
+ hidden_channels,
277
+ filter_channels,
278
+ n_heads,
279
+ n_layers,
280
+ kernel_size,
281
+ p_dropout)
282
+ self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
283
+ self.f0_prenet = nn.Conv1d(1, hidden_channels , 3, padding=1)
284
+ self.cond = nn.Conv1d(spk_channels, hidden_channels, 1)
285
+
286
+ def forward(self, x, norm_f0, x_mask, spk_emb=None):
287
+ x = torch.detach(x)
288
+ if (spk_emb is not None):
289
+ x = x + self.cond(spk_emb)
290
+ x += self.f0_prenet(norm_f0)
291
+ x = self.prenet(x) * x_mask
292
+ x = self.decoder(x * x_mask, x_mask)
293
+ x = self.proj(x) * x_mask
294
+ return x
295
+
296
+
297
+ class SynthesizerTrn(nn.Module):
298
+ """
299
+ Synthesizer for Training
300
+ """
301
+
302
+ def __init__(self,
303
+ spec_channels,
304
+ segment_size,
305
+ inter_channels,
306
+ hidden_channels,
307
+ filter_channels,
308
+ n_heads,
309
+ n_layers,
310
+ kernel_size,
311
+ p_dropout,
312
+ resblock,
313
+ resblock_kernel_sizes,
314
+ resblock_dilation_sizes,
315
+ upsample_rates,
316
+ upsample_initial_channel,
317
+ upsample_kernel_sizes,
318
+ gin_channels,
319
+ ssl_dim,
320
+ n_speakers,
321
+ sampling_rate=44100,
322
+ **kwargs):
323
+
324
+ super().__init__()
325
+ self.spec_channels = spec_channels
326
+ self.inter_channels = inter_channels
327
+ self.hidden_channels = hidden_channels
328
+ self.filter_channels = filter_channels
329
+ self.n_heads = n_heads
330
+ self.n_layers = n_layers
331
+ self.kernel_size = kernel_size
332
+ self.p_dropout = p_dropout
333
+ self.resblock = resblock
334
+ self.resblock_kernel_sizes = resblock_kernel_sizes
335
+ self.resblock_dilation_sizes = resblock_dilation_sizes
336
+ self.upsample_rates = upsample_rates
337
+ self.upsample_initial_channel = upsample_initial_channel
338
+ self.upsample_kernel_sizes = upsample_kernel_sizes
339
+ self.segment_size = segment_size
340
+ self.gin_channels = gin_channels
341
+ self.ssl_dim = ssl_dim
342
+ self.emb_g = nn.Embedding(n_speakers, gin_channels)
343
+
344
+ self.pre = nn.Conv1d(ssl_dim, hidden_channels, kernel_size=5, padding=2)
345
+
346
+ self.enc_p = TextEncoder(
347
+ inter_channels,
348
+ hidden_channels,
349
+ filter_channels=filter_channels,
350
+ n_heads=n_heads,
351
+ n_layers=n_layers,
352
+ kernel_size=kernel_size,
353
+ p_dropout=p_dropout
354
+ )
355
+ hps = {
356
+ "sampling_rate": sampling_rate,
357
+ "inter_channels": inter_channels,
358
+ "resblock": resblock,
359
+ "resblock_kernel_sizes": resblock_kernel_sizes,
360
+ "resblock_dilation_sizes": resblock_dilation_sizes,
361
+ "upsample_rates": upsample_rates,
362
+ "upsample_initial_channel": upsample_initial_channel,
363
+ "upsample_kernel_sizes": upsample_kernel_sizes,
364
+ "gin_channels": gin_channels,
365
+ }
366
+ self.dec = Generator(h=hps)
367
+ self.enc_q = Encoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
368
+ self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
369
+ self.f0_decoder = F0Decoder(
370
+ 1,
371
+ hidden_channels,
372
+ filter_channels,
373
+ n_heads,
374
+ n_layers,
375
+ kernel_size,
376
+ p_dropout,
377
+ spk_channels=gin_channels
378
+ )
379
+ self.emb_uv = nn.Embedding(2, hidden_channels)
380
+
381
+ def forward(self, c, f0, uv, spec, g=None, c_lengths=None, spec_lengths=None):
382
+ g = self.emb_g(g).transpose(1,2)
383
+ # ssl prenet
384
+ x_mask = torch.unsqueeze(commons.sequence_mask(c_lengths, c.size(2)), 1).to(c.dtype)
385
+ x = self.pre(c) * x_mask + self.emb_uv(uv.long()).transpose(1,2)
386
+
387
+ # f0 predict
388
+ lf0 = 2595. * torch.log10(1. + f0.unsqueeze(1) / 700.) / 500
389
+ norm_lf0 = utils.normalize_f0(lf0, x_mask, uv)
390
+ pred_lf0 = self.f0_decoder(x, norm_lf0, x_mask, spk_emb=g)
391
+
392
+ # encoder
393
+ z_ptemp, m_p, logs_p, _ = self.enc_p(x, x_mask, f0=f0_to_coarse(f0))
394
+ z, m_q, logs_q, spec_mask = self.enc_q(spec, spec_lengths, g=g)
395
+
396
+ # flow
397
+ z_p = self.flow(z, spec_mask, g=g)
398
+ z_slice, pitch_slice, ids_slice = commons.rand_slice_segments_with_pitch(z, f0, spec_lengths, self.segment_size)
399
+
400
+ # nsf decoder
401
+ o = self.dec(z_slice, g=g, f0=pitch_slice)
402
+
403
+ return o, ids_slice, spec_mask, (z, z_p, m_p, logs_p, m_q, logs_q), pred_lf0, norm_lf0, lf0
404
+
405
+ def infer(self, c, f0, uv, g=None, noice_scale=0.35, predict_f0=False):
406
+ c_lengths = (torch.ones(c.size(0)) * c.size(-1)).to(c.device)
407
+ g = self.emb_g(g).transpose(1,2)
408
+ x_mask = torch.unsqueeze(commons.sequence_mask(c_lengths, c.size(2)), 1).to(c.dtype)
409
+ x = self.pre(c) * x_mask + self.emb_uv(uv.long()).transpose(1,2)
410
+
411
+ if predict_f0:
412
+ lf0 = 2595. * torch.log10(1. + f0.unsqueeze(1) / 700.) / 500
413
+ norm_lf0 = utils.normalize_f0(lf0, x_mask, uv, random_scale=False)
414
+ pred_lf0 = self.f0_decoder(x, norm_lf0, x_mask, spk_emb=g)
415
+ f0 = (700 * (torch.pow(10, pred_lf0 * 500 / 2595) - 1)).squeeze(1)
416
+
417
+ z_p, m_p, logs_p, c_mask = self.enc_p(x, x_mask, f0=f0_to_coarse(f0), noice_scale=noice_scale)
418
+ z = self.flow(z_p, c_mask, g=g, reverse=True)
419
+ o = self.dec(z * c_mask, g=g, f0=f0)
420
+ return o
models/tannhauser/config.json ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "train": {
3
+ "log_interval": 200,
4
+ "eval_interval": 800,
5
+ "seed": 1234,
6
+ "epochs": 10000,
7
+ "learning_rate": 0.0001,
8
+ "betas": [
9
+ 0.8,
10
+ 0.99
11
+ ],
12
+ "eps": 1e-09,
13
+ "batch_size": 32,
14
+ "fp16_run": false,
15
+ "lr_decay": 0.999875,
16
+ "segment_size": 10240,
17
+ "init_lr_ratio": 1,
18
+ "warmup_epochs": 0,
19
+ "c_mel": 45,
20
+ "c_kl": 1.0,
21
+ "use_sr": true,
22
+ "max_speclen": 512,
23
+ "port": "8001",
24
+ "keep_ckpts": 99
25
+ },
26
+ "data": {
27
+ "training_files": "filelists/train.txt",
28
+ "validation_files": "filelists/val.txt",
29
+ "max_wav_value": 32768.0,
30
+ "sampling_rate": 44100,
31
+ "filter_length": 2048,
32
+ "hop_length": 512,
33
+ "win_length": 2048,
34
+ "n_mel_channels": 80,
35
+ "mel_fmin": 0.0,
36
+ "mel_fmax": 22050
37
+ },
38
+ "model": {
39
+ "inter_channels": 192,
40
+ "hidden_channels": 192,
41
+ "filter_channels": 768,
42
+ "n_heads": 2,
43
+ "n_layers": 6,
44
+ "kernel_size": 3,
45
+ "p_dropout": 0.1,
46
+ "resblock": "1",
47
+ "resblock_kernel_sizes": [
48
+ 3,
49
+ 7,
50
+ 11
51
+ ],
52
+ "resblock_dilation_sizes": [
53
+ [
54
+ 1,
55
+ 3,
56
+ 5
57
+ ],
58
+ [
59
+ 1,
60
+ 3,
61
+ 5
62
+ ],
63
+ [
64
+ 1,
65
+ 3,
66
+ 5
67
+ ]
68
+ ],
69
+ "upsample_rates": [
70
+ 8,
71
+ 8,
72
+ 2,
73
+ 2,
74
+ 2
75
+ ],
76
+ "upsample_initial_channel": 512,
77
+ "upsample_kernel_sizes": [
78
+ 16,
79
+ 16,
80
+ 4,
81
+ 4,
82
+ 4
83
+ ],
84
+ "n_layers_q": 3,
85
+ "use_spectral_norm": false,
86
+ "gin_channels": 256,
87
+ "ssl_dim": 256,
88
+ "n_speakers": 200
89
+ },
90
+ "spk": {
91
+ "tannhauser": 0
92
+ }
93
+ }
models/tannhauser/tannhauser.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a39d57f9e5ae1eba070eb782df6355699ff93f680b075f99c45613ad590035ef
3
+ size 180883747
models/teio/config.json ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "train": {
3
+ "log_interval": 200,
4
+ "eval_interval": 800,
5
+ "seed": 1234,
6
+ "epochs": 10000,
7
+ "learning_rate": 0.0001,
8
+ "betas": [
9
+ 0.8,
10
+ 0.99
11
+ ],
12
+ "eps": 1e-09,
13
+ "batch_size": 32,
14
+ "fp16_run": false,
15
+ "lr_decay": 0.999875,
16
+ "segment_size": 10240,
17
+ "init_lr_ratio": 1,
18
+ "warmup_epochs": 0,
19
+ "c_mel": 45,
20
+ "c_kl": 1.0,
21
+ "use_sr": true,
22
+ "max_speclen": 512,
23
+ "port": "8001",
24
+ "keep_ckpts": 99
25
+ },
26
+ "data": {
27
+ "training_files": "filelists/train.txt",
28
+ "validation_files": "filelists/val.txt",
29
+ "max_wav_value": 32768.0,
30
+ "sampling_rate": 44100,
31
+ "filter_length": 2048,
32
+ "hop_length": 512,
33
+ "win_length": 2048,
34
+ "n_mel_channels": 80,
35
+ "mel_fmin": 0.0,
36
+ "mel_fmax": 22050
37
+ },
38
+ "model": {
39
+ "inter_channels": 192,
40
+ "hidden_channels": 192,
41
+ "filter_channels": 768,
42
+ "n_heads": 2,
43
+ "n_layers": 6,
44
+ "kernel_size": 3,
45
+ "p_dropout": 0.1,
46
+ "resblock": "1",
47
+ "resblock_kernel_sizes": [
48
+ 3,
49
+ 7,
50
+ 11
51
+ ],
52
+ "resblock_dilation_sizes": [
53
+ [
54
+ 1,
55
+ 3,
56
+ 5
57
+ ],
58
+ [
59
+ 1,
60
+ 3,
61
+ 5
62
+ ],
63
+ [
64
+ 1,
65
+ 3,
66
+ 5
67
+ ]
68
+ ],
69
+ "upsample_rates": [
70
+ 8,
71
+ 8,
72
+ 2,
73
+ 2,
74
+ 2
75
+ ],
76
+ "upsample_initial_channel": 512,
77
+ "upsample_kernel_sizes": [
78
+ 16,
79
+ 16,
80
+ 4,
81
+ 4,
82
+ 4
83
+ ],
84
+ "n_layers_q": 3,
85
+ "use_spectral_norm": false,
86
+ "gin_channels": 256,
87
+ "ssl_dim": 256,
88
+ "n_speakers": 200
89
+ },
90
+ "spk": {
91
+ "teio": 0
92
+ }
93
+ }
models/teio/teio.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:305cbd8bc9bc468f2744d0fc425d1c7363a6232140728d403e90486dc2921160
3
+ size 180883747
modules/__init__.py ADDED
File without changes
modules/__pycache__/__init__.cpython-38.pyc ADDED
Binary file (135 Bytes). View file
 
modules/__pycache__/attentions.cpython-38.pyc ADDED
Binary file (10.6 kB). View file
 
modules/__pycache__/commons.cpython-38.pyc ADDED
Binary file (6.63 kB). View file
 
modules/__pycache__/modules.cpython-38.pyc ADDED
Binary file (10.1 kB). View file
 
modules/attentions.py ADDED
@@ -0,0 +1,349 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import math
3
+ import numpy as np
4
+ import torch
5
+ from torch import nn
6
+ from torch.nn import functional as F
7
+
8
+ import modules.commons as commons
9
+ import modules.modules as modules
10
+ from modules.modules import LayerNorm
11
+
12
+
13
+ class FFT(nn.Module):
14
+ def __init__(self, hidden_channels, filter_channels, n_heads, n_layers=1, kernel_size=1, p_dropout=0.,
15
+ proximal_bias=False, proximal_init=True, **kwargs):
16
+ super().__init__()
17
+ self.hidden_channels = hidden_channels
18
+ self.filter_channels = filter_channels
19
+ self.n_heads = n_heads
20
+ self.n_layers = n_layers
21
+ self.kernel_size = kernel_size
22
+ self.p_dropout = p_dropout
23
+ self.proximal_bias = proximal_bias
24
+ self.proximal_init = proximal_init
25
+
26
+ self.drop = nn.Dropout(p_dropout)
27
+ self.self_attn_layers = nn.ModuleList()
28
+ self.norm_layers_0 = nn.ModuleList()
29
+ self.ffn_layers = nn.ModuleList()
30
+ self.norm_layers_1 = nn.ModuleList()
31
+ for i in range(self.n_layers):
32
+ self.self_attn_layers.append(
33
+ MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias,
34
+ proximal_init=proximal_init))
35
+ self.norm_layers_0.append(LayerNorm(hidden_channels))
36
+ self.ffn_layers.append(
37
+ FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
38
+ self.norm_layers_1.append(LayerNorm(hidden_channels))
39
+
40
+ def forward(self, x, x_mask):
41
+ """
42
+ x: decoder input
43
+ h: encoder output
44
+ """
45
+ self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)
46
+ x = x * x_mask
47
+ for i in range(self.n_layers):
48
+ y = self.self_attn_layers[i](x, x, self_attn_mask)
49
+ y = self.drop(y)
50
+ x = self.norm_layers_0[i](x + y)
51
+
52
+ y = self.ffn_layers[i](x, x_mask)
53
+ y = self.drop(y)
54
+ x = self.norm_layers_1[i](x + y)
55
+ x = x * x_mask
56
+ return x
57
+
58
+
59
+ class Encoder(nn.Module):
60
+ def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, **kwargs):
61
+ super().__init__()
62
+ self.hidden_channels = hidden_channels
63
+ self.filter_channels = filter_channels
64
+ self.n_heads = n_heads
65
+ self.n_layers = n_layers
66
+ self.kernel_size = kernel_size
67
+ self.p_dropout = p_dropout
68
+ self.window_size = window_size
69
+
70
+ self.drop = nn.Dropout(p_dropout)
71
+ self.attn_layers = nn.ModuleList()
72
+ self.norm_layers_1 = nn.ModuleList()
73
+ self.ffn_layers = nn.ModuleList()
74
+ self.norm_layers_2 = nn.ModuleList()
75
+ for i in range(self.n_layers):
76
+ self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))
77
+ self.norm_layers_1.append(LayerNorm(hidden_channels))
78
+ self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))
79
+ self.norm_layers_2.append(LayerNorm(hidden_channels))
80
+
81
+ def forward(self, x, x_mask):
82
+ attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
83
+ x = x * x_mask
84
+ for i in range(self.n_layers):
85
+ y = self.attn_layers[i](x, x, attn_mask)
86
+ y = self.drop(y)
87
+ x = self.norm_layers_1[i](x + y)
88
+
89
+ y = self.ffn_layers[i](x, x_mask)
90
+ y = self.drop(y)
91
+ x = self.norm_layers_2[i](x + y)
92
+ x = x * x_mask
93
+ return x
94
+
95
+
96
+ class Decoder(nn.Module):
97
+ def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs):
98
+ super().__init__()
99
+ self.hidden_channels = hidden_channels
100
+ self.filter_channels = filter_channels
101
+ self.n_heads = n_heads
102
+ self.n_layers = n_layers
103
+ self.kernel_size = kernel_size
104
+ self.p_dropout = p_dropout
105
+ self.proximal_bias = proximal_bias
106
+ self.proximal_init = proximal_init
107
+
108
+ self.drop = nn.Dropout(p_dropout)
109
+ self.self_attn_layers = nn.ModuleList()
110
+ self.norm_layers_0 = nn.ModuleList()
111
+ self.encdec_attn_layers = nn.ModuleList()
112
+ self.norm_layers_1 = nn.ModuleList()
113
+ self.ffn_layers = nn.ModuleList()
114
+ self.norm_layers_2 = nn.ModuleList()
115
+ for i in range(self.n_layers):
116
+ self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))
117
+ self.norm_layers_0.append(LayerNorm(hidden_channels))
118
+ self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))
119
+ self.norm_layers_1.append(LayerNorm(hidden_channels))
120
+ self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
121
+ self.norm_layers_2.append(LayerNorm(hidden_channels))
122
+
123
+ def forward(self, x, x_mask, h, h_mask):
124
+ """
125
+ x: decoder input
126
+ h: encoder output
127
+ """
128
+ self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)
129
+ encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
130
+ x = x * x_mask
131
+ for i in range(self.n_layers):
132
+ y = self.self_attn_layers[i](x, x, self_attn_mask)
133
+ y = self.drop(y)
134
+ x = self.norm_layers_0[i](x + y)
135
+
136
+ y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
137
+ y = self.drop(y)
138
+ x = self.norm_layers_1[i](x + y)
139
+
140
+ y = self.ffn_layers[i](x, x_mask)
141
+ y = self.drop(y)
142
+ x = self.norm_layers_2[i](x + y)
143
+ x = x * x_mask
144
+ return x
145
+
146
+
147
+ class MultiHeadAttention(nn.Module):
148
+ def __init__(self, channels, out_channels, n_heads, p_dropout=0., window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False):
149
+ super().__init__()
150
+ assert channels % n_heads == 0
151
+
152
+ self.channels = channels
153
+ self.out_channels = out_channels
154
+ self.n_heads = n_heads
155
+ self.p_dropout = p_dropout
156
+ self.window_size = window_size
157
+ self.heads_share = heads_share
158
+ self.block_length = block_length
159
+ self.proximal_bias = proximal_bias
160
+ self.proximal_init = proximal_init
161
+ self.attn = None
162
+
163
+ self.k_channels = channels // n_heads
164
+ self.conv_q = nn.Conv1d(channels, channels, 1)
165
+ self.conv_k = nn.Conv1d(channels, channels, 1)
166
+ self.conv_v = nn.Conv1d(channels, channels, 1)
167
+ self.conv_o = nn.Conv1d(channels, out_channels, 1)
168
+ self.drop = nn.Dropout(p_dropout)
169
+
170
+ if window_size is not None:
171
+ n_heads_rel = 1 if heads_share else n_heads
172
+ rel_stddev = self.k_channels**-0.5
173
+ self.emb_rel_k = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
174
+ self.emb_rel_v = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
175
+
176
+ nn.init.xavier_uniform_(self.conv_q.weight)
177
+ nn.init.xavier_uniform_(self.conv_k.weight)
178
+ nn.init.xavier_uniform_(self.conv_v.weight)
179
+ if proximal_init:
180
+ with torch.no_grad():
181
+ self.conv_k.weight.copy_(self.conv_q.weight)
182
+ self.conv_k.bias.copy_(self.conv_q.bias)
183
+
184
+ def forward(self, x, c, attn_mask=None):
185
+ q = self.conv_q(x)
186
+ k = self.conv_k(c)
187
+ v = self.conv_v(c)
188
+
189
+ x, self.attn = self.attention(q, k, v, mask=attn_mask)
190
+
191
+ x = self.conv_o(x)
192
+ return x
193
+
194
+ def attention(self, query, key, value, mask=None):
195
+ # reshape [b, d, t] -> [b, n_h, t, d_k]
196
+ b, d, t_s, t_t = (*key.size(), query.size(2))
197
+ query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
198
+ key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
199
+ value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
200
+
201
+ scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
202
+ if self.window_size is not None:
203
+ assert t_s == t_t, "Relative attention is only available for self-attention."
204
+ key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
205
+ rel_logits = self._matmul_with_relative_keys(query /math.sqrt(self.k_channels), key_relative_embeddings)
206
+ scores_local = self._relative_position_to_absolute_position(rel_logits)
207
+ scores = scores + scores_local
208
+ if self.proximal_bias:
209
+ assert t_s == t_t, "Proximal bias is only available for self-attention."
210
+ scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)
211
+ if mask is not None:
212
+ scores = scores.masked_fill(mask == 0, -1e4)
213
+ if self.block_length is not None:
214
+ assert t_s == t_t, "Local attention is only available for self-attention."
215
+ block_mask = torch.ones_like(scores).triu(-self.block_length).tril(self.block_length)
216
+ scores = scores.masked_fill(block_mask == 0, -1e4)
217
+ p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]
218
+ p_attn = self.drop(p_attn)
219
+ output = torch.matmul(p_attn, value)
220
+ if self.window_size is not None:
221
+ relative_weights = self._absolute_position_to_relative_position(p_attn)
222
+ value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
223
+ output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
224
+ output = output.transpose(2, 3).contiguous().view(b, d, t_t) # [b, n_h, t_t, d_k] -> [b, d, t_t]
225
+ return output, p_attn
226
+
227
+ def _matmul_with_relative_values(self, x, y):
228
+ """
229
+ x: [b, h, l, m]
230
+ y: [h or 1, m, d]
231
+ ret: [b, h, l, d]
232
+ """
233
+ ret = torch.matmul(x, y.unsqueeze(0))
234
+ return ret
235
+
236
+ def _matmul_with_relative_keys(self, x, y):
237
+ """
238
+ x: [b, h, l, d]
239
+ y: [h or 1, m, d]
240
+ ret: [b, h, l, m]
241
+ """
242
+ ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
243
+ return ret
244
+
245
+ def _get_relative_embeddings(self, relative_embeddings, length):
246
+ max_relative_position = 2 * self.window_size + 1
247
+ # Pad first before slice to avoid using cond ops.
248
+ pad_length = max(length - (self.window_size + 1), 0)
249
+ slice_start_position = max((self.window_size + 1) - length, 0)
250
+ slice_end_position = slice_start_position + 2 * length - 1
251
+ if pad_length > 0:
252
+ padded_relative_embeddings = F.pad(
253
+ relative_embeddings,
254
+ commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))
255
+ else:
256
+ padded_relative_embeddings = relative_embeddings
257
+ used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position]
258
+ return used_relative_embeddings
259
+
260
+ def _relative_position_to_absolute_position(self, x):
261
+ """
262
+ x: [b, h, l, 2*l-1]
263
+ ret: [b, h, l, l]
264
+ """
265
+ batch, heads, length, _ = x.size()
266
+ # Concat columns of pad to shift from relative to absolute indexing.
267
+ x = F.pad(x, commons.convert_pad_shape([[0,0],[0,0],[0,0],[0,1]]))
268
+
269
+ # Concat extra elements so to add up to shape (len+1, 2*len-1).
270
+ x_flat = x.view([batch, heads, length * 2 * length])
271
+ x_flat = F.pad(x_flat, commons.convert_pad_shape([[0,0],[0,0],[0,length-1]]))
272
+
273
+ # Reshape and slice out the padded elements.
274
+ x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]
275
+ return x_final
276
+
277
+ def _absolute_position_to_relative_position(self, x):
278
+ """
279
+ x: [b, h, l, l]
280
+ ret: [b, h, l, 2*l-1]
281
+ """
282
+ batch, heads, length, _ = x.size()
283
+ # padd along column
284
+ x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]]))
285
+ x_flat = x.view([batch, heads, length**2 + length*(length -1)])
286
+ # add 0's in the beginning that will skew the elements after reshape
287
+ x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
288
+ x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:]
289
+ return x_final
290
+
291
+ def _attention_bias_proximal(self, length):
292
+ """Bias for self-attention to encourage attention to close positions.
293
+ Args:
294
+ length: an integer scalar.
295
+ Returns:
296
+ a Tensor with shape [1, 1, length, length]
297
+ """
298
+ r = torch.arange(length, dtype=torch.float32)
299
+ diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
300
+ return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
301
+
302
+
303
+ class FFN(nn.Module):
304
+ def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False):
305
+ super().__init__()
306
+ self.in_channels = in_channels
307
+ self.out_channels = out_channels
308
+ self.filter_channels = filter_channels
309
+ self.kernel_size = kernel_size
310
+ self.p_dropout = p_dropout
311
+ self.activation = activation
312
+ self.causal = causal
313
+
314
+ if causal:
315
+ self.padding = self._causal_padding
316
+ else:
317
+ self.padding = self._same_padding
318
+
319
+ self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
320
+ self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
321
+ self.drop = nn.Dropout(p_dropout)
322
+
323
+ def forward(self, x, x_mask):
324
+ x = self.conv_1(self.padding(x * x_mask))
325
+ if self.activation == "gelu":
326
+ x = x * torch.sigmoid(1.702 * x)
327
+ else:
328
+ x = torch.relu(x)
329
+ x = self.drop(x)
330
+ x = self.conv_2(self.padding(x * x_mask))
331
+ return x * x_mask
332
+
333
+ def _causal_padding(self, x):
334
+ if self.kernel_size == 1:
335
+ return x
336
+ pad_l = self.kernel_size - 1
337
+ pad_r = 0
338
+ padding = [[0, 0], [0, 0], [pad_l, pad_r]]
339
+ x = F.pad(x, commons.convert_pad_shape(padding))
340
+ return x
341
+
342
+ def _same_padding(self, x):
343
+ if self.kernel_size == 1:
344
+ return x
345
+ pad_l = (self.kernel_size - 1) // 2
346
+ pad_r = self.kernel_size // 2
347
+ padding = [[0, 0], [0, 0], [pad_l, pad_r]]
348
+ x = F.pad(x, commons.convert_pad_shape(padding))
349
+ return x
modules/commons.py ADDED
@@ -0,0 +1,188 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import numpy as np
3
+ import torch
4
+ from torch import nn
5
+ from torch.nn import functional as F
6
+
7
+ def slice_pitch_segments(x, ids_str, segment_size=4):
8
+ ret = torch.zeros_like(x[:, :segment_size])
9
+ for i in range(x.size(0)):
10
+ idx_str = ids_str[i]
11
+ idx_end = idx_str + segment_size
12
+ ret[i] = x[i, idx_str:idx_end]
13
+ return ret
14
+
15
+ def rand_slice_segments_with_pitch(x, pitch, x_lengths=None, segment_size=4):
16
+ b, d, t = x.size()
17
+ if x_lengths is None:
18
+ x_lengths = t
19
+ ids_str_max = x_lengths - segment_size + 1
20
+ ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
21
+ ret = slice_segments(x, ids_str, segment_size)
22
+ ret_pitch = slice_pitch_segments(pitch, ids_str, segment_size)
23
+ return ret, ret_pitch, ids_str
24
+
25
+ def init_weights(m, mean=0.0, std=0.01):
26
+ classname = m.__class__.__name__
27
+ if classname.find("Conv") != -1:
28
+ m.weight.data.normal_(mean, std)
29
+
30
+
31
+ def get_padding(kernel_size, dilation=1):
32
+ return int((kernel_size*dilation - dilation)/2)
33
+
34
+
35
+ def convert_pad_shape(pad_shape):
36
+ l = pad_shape[::-1]
37
+ pad_shape = [item for sublist in l for item in sublist]
38
+ return pad_shape
39
+
40
+
41
+ def intersperse(lst, item):
42
+ result = [item] * (len(lst) * 2 + 1)
43
+ result[1::2] = lst
44
+ return result
45
+
46
+
47
+ def kl_divergence(m_p, logs_p, m_q, logs_q):
48
+ """KL(P||Q)"""
49
+ kl = (logs_q - logs_p) - 0.5
50
+ kl += 0.5 * (torch.exp(2. * logs_p) + ((m_p - m_q)**2)) * torch.exp(-2. * logs_q)
51
+ return kl
52
+
53
+
54
+ def rand_gumbel(shape):
55
+ """Sample from the Gumbel distribution, protect from overflows."""
56
+ uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
57
+ return -torch.log(-torch.log(uniform_samples))
58
+
59
+
60
+ def rand_gumbel_like(x):
61
+ g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
62
+ return g
63
+
64
+
65
+ def slice_segments(x, ids_str, segment_size=4):
66
+ ret = torch.zeros_like(x[:, :, :segment_size])
67
+ for i in range(x.size(0)):
68
+ idx_str = ids_str[i]
69
+ idx_end = idx_str + segment_size
70
+ ret[i] = x[i, :, idx_str:idx_end]
71
+ return ret
72
+
73
+
74
+ def rand_slice_segments(x, x_lengths=None, segment_size=4):
75
+ b, d, t = x.size()
76
+ if x_lengths is None:
77
+ x_lengths = t
78
+ ids_str_max = x_lengths - segment_size + 1
79
+ ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
80
+ ret = slice_segments(x, ids_str, segment_size)
81
+ return ret, ids_str
82
+
83
+
84
+ def rand_spec_segments(x, x_lengths=None, segment_size=4):
85
+ b, d, t = x.size()
86
+ if x_lengths is None:
87
+ x_lengths = t
88
+ ids_str_max = x_lengths - segment_size
89
+ ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
90
+ ret = slice_segments(x, ids_str, segment_size)
91
+ return ret, ids_str
92
+
93
+
94
+ def get_timing_signal_1d(
95
+ length, channels, min_timescale=1.0, max_timescale=1.0e4):
96
+ position = torch.arange(length, dtype=torch.float)
97
+ num_timescales = channels // 2
98
+ log_timescale_increment = (
99
+ math.log(float(max_timescale) / float(min_timescale)) /
100
+ (num_timescales - 1))
101
+ inv_timescales = min_timescale * torch.exp(
102
+ torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment)
103
+ scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
104
+ signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
105
+ signal = F.pad(signal, [0, 0, 0, channels % 2])
106
+ signal = signal.view(1, channels, length)
107
+ return signal
108
+
109
+
110
+ def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
111
+ b, channels, length = x.size()
112
+ signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
113
+ return x + signal.to(dtype=x.dtype, device=x.device)
114
+
115
+
116
+ def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
117
+ b, channels, length = x.size()
118
+ signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
119
+ return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
120
+
121
+
122
+ def subsequent_mask(length):
123
+ mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
124
+ return mask
125
+
126
+
127
+ @torch.jit.script
128
+ def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
129
+ n_channels_int = n_channels[0]
130
+ in_act = input_a + input_b
131
+ t_act = torch.tanh(in_act[:, :n_channels_int, :])
132
+ s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
133
+ acts = t_act * s_act
134
+ return acts
135
+
136
+
137
+ def convert_pad_shape(pad_shape):
138
+ l = pad_shape[::-1]
139
+ pad_shape = [item for sublist in l for item in sublist]
140
+ return pad_shape
141
+
142
+
143
+ def shift_1d(x):
144
+ x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
145
+ return x
146
+
147
+
148
+ def sequence_mask(length, max_length=None):
149
+ if max_length is None:
150
+ max_length = length.max()
151
+ x = torch.arange(max_length, dtype=length.dtype, device=length.device)
152
+ return x.unsqueeze(0) < length.unsqueeze(1)
153
+
154
+
155
+ def generate_path(duration, mask):
156
+ """
157
+ duration: [b, 1, t_x]
158
+ mask: [b, 1, t_y, t_x]
159
+ """
160
+ device = duration.device
161
+
162
+ b, _, t_y, t_x = mask.shape
163
+ cum_duration = torch.cumsum(duration, -1)
164
+
165
+ cum_duration_flat = cum_duration.view(b * t_x)
166
+ path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
167
+ path = path.view(b, t_x, t_y)
168
+ path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
169
+ path = path.unsqueeze(1).transpose(2,3) * mask
170
+ return path
171
+
172
+
173
+ def clip_grad_value_(parameters, clip_value, norm_type=2):
174
+ if isinstance(parameters, torch.Tensor):
175
+ parameters = [parameters]
176
+ parameters = list(filter(lambda p: p.grad is not None, parameters))
177
+ norm_type = float(norm_type)
178
+ if clip_value is not None:
179
+ clip_value = float(clip_value)
180
+
181
+ total_norm = 0
182
+ for p in parameters:
183
+ param_norm = p.grad.data.norm(norm_type)
184
+ total_norm += param_norm.item() ** norm_type
185
+ if clip_value is not None:
186
+ p.grad.data.clamp_(min=-clip_value, max=clip_value)
187
+ total_norm = total_norm ** (1. / norm_type)
188
+ return total_norm
modules/ddsp.py ADDED
@@ -0,0 +1,190 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+ from torch.nn import functional as F
4
+ import torch.fft as fft
5
+ import numpy as np
6
+ import librosa as li
7
+ import math
8
+ from scipy.signal import get_window
9
+
10
+
11
+ def safe_log(x):
12
+ return torch.log(x + 1e-7)
13
+
14
+
15
+ @torch.no_grad()
16
+ def mean_std_loudness(dataset):
17
+ mean = 0
18
+ std = 0
19
+ n = 0
20
+ for _, _, l in dataset:
21
+ n += 1
22
+ mean += (l.mean().item() - mean) / n
23
+ std += (l.std().item() - std) / n
24
+ return mean, std
25
+
26
+
27
+ def multiscale_fft(signal, scales, overlap):
28
+ stfts = []
29
+ for s in scales:
30
+ S = torch.stft(
31
+ signal,
32
+ s,
33
+ int(s * (1 - overlap)),
34
+ s,
35
+ torch.hann_window(s).to(signal),
36
+ True,
37
+ normalized=True,
38
+ return_complex=True,
39
+ ).abs()
40
+ stfts.append(S)
41
+ return stfts
42
+
43
+
44
+ def resample(x, factor: int):
45
+ batch, frame, channel = x.shape
46
+ x = x.permute(0, 2, 1).reshape(batch * channel, 1, frame)
47
+
48
+ window = torch.hann_window(
49
+ factor * 2,
50
+ dtype=x.dtype,
51
+ device=x.device,
52
+ ).reshape(1, 1, -1)
53
+ y = torch.zeros(x.shape[0], x.shape[1], factor * x.shape[2]).to(x)
54
+ y[..., ::factor] = x
55
+ y[..., -1:] = x[..., -1:]
56
+ y = torch.nn.functional.pad(y, [factor, factor])
57
+ y = torch.nn.functional.conv1d(y, window)[..., :-1]
58
+
59
+ y = y.reshape(batch, channel, factor * frame).permute(0, 2, 1)
60
+
61
+ return y
62
+
63
+
64
+ def upsample(signal, factor):
65
+ signal = signal.permute(0, 2, 1)
66
+ signal = nn.functional.interpolate(signal, size=signal.shape[-1] * factor)
67
+ return signal.permute(0, 2, 1)
68
+
69
+
70
+ def remove_above_nyquist(amplitudes, pitch, sampling_rate):
71
+ n_harm = amplitudes.shape[-1]
72
+ pitches = pitch * torch.arange(1, n_harm + 1).to(pitch)
73
+ aa = (pitches < sampling_rate / 2).float() + 1e-4
74
+ return amplitudes * aa
75
+
76
+
77
+ def scale_function(x):
78
+ return 2 * torch.sigmoid(x) ** (math.log(10)) + 1e-7
79
+
80
+
81
+ def extract_loudness(signal, sampling_rate, block_size, n_fft=2048):
82
+ S = li.stft(
83
+ signal,
84
+ n_fft=n_fft,
85
+ hop_length=block_size,
86
+ win_length=n_fft,
87
+ center=True,
88
+ )
89
+ S = np.log(abs(S) + 1e-7)
90
+ f = li.fft_frequencies(sampling_rate, n_fft)
91
+ a_weight = li.A_weighting(f)
92
+
93
+ S = S + a_weight.reshape(-1, 1)
94
+
95
+ S = np.mean(S, 0)[..., :-1]
96
+
97
+ return S
98
+
99
+
100
+ def extract_pitch(signal, sampling_rate, block_size):
101
+ length = signal.shape[-1] // block_size
102
+ f0 = crepe.predict(
103
+ signal,
104
+ sampling_rate,
105
+ step_size=int(1000 * block_size / sampling_rate),
106
+ verbose=1,
107
+ center=True,
108
+ viterbi=True,
109
+ )
110
+ f0 = f0[1].reshape(-1)[:-1]
111
+
112
+ if f0.shape[-1] != length:
113
+ f0 = np.interp(
114
+ np.linspace(0, 1, length, endpoint=False),
115
+ np.linspace(0, 1, f0.shape[-1], endpoint=False),
116
+ f0,
117
+ )
118
+
119
+ return f0
120
+
121
+
122
+ def mlp(in_size, hidden_size, n_layers):
123
+ channels = [in_size] + (n_layers) * [hidden_size]
124
+ net = []
125
+ for i in range(n_layers):
126
+ net.append(nn.Linear(channels[i], channels[i + 1]))
127
+ net.append(nn.LayerNorm(channels[i + 1]))
128
+ net.append(nn.LeakyReLU())
129
+ return nn.Sequential(*net)
130
+
131
+
132
+ def gru(n_input, hidden_size):
133
+ return nn.GRU(n_input * hidden_size, hidden_size, batch_first=True)
134
+
135
+
136
+ def harmonic_synth(pitch, amplitudes, sampling_rate):
137
+ n_harmonic = amplitudes.shape[-1]
138
+ omega = torch.cumsum(2 * math.pi * pitch / sampling_rate, 1)
139
+ omegas = omega * torch.arange(1, n_harmonic + 1).to(omega)
140
+ signal = (torch.sin(omegas) * amplitudes).sum(-1, keepdim=True)
141
+ return signal
142
+
143
+
144
+ def amp_to_impulse_response(amp, target_size):
145
+ amp = torch.stack([amp, torch.zeros_like(amp)], -1)
146
+ amp = torch.view_as_complex(amp)
147
+ amp = fft.irfft(amp)
148
+
149
+ filter_size = amp.shape[-1]
150
+
151
+ amp = torch.roll(amp, filter_size // 2, -1)
152
+ win = torch.hann_window(filter_size, dtype=amp.dtype, device=amp.device)
153
+
154
+ amp = amp * win
155
+
156
+ amp = nn.functional.pad(amp, (0, int(target_size) - int(filter_size)))
157
+ amp = torch.roll(amp, -filter_size // 2, -1)
158
+
159
+ return amp
160
+
161
+
162
+ def fft_convolve(signal, kernel):
163
+ signal = nn.functional.pad(signal, (0, signal.shape[-1]))
164
+ kernel = nn.functional.pad(kernel, (kernel.shape[-1], 0))
165
+
166
+ output = fft.irfft(fft.rfft(signal) * fft.rfft(kernel))
167
+ output = output[..., output.shape[-1] // 2:]
168
+
169
+ return output
170
+
171
+
172
+ def init_kernels(win_len, win_inc, fft_len, win_type=None, invers=False):
173
+ if win_type == 'None' or win_type is None:
174
+ window = np.ones(win_len)
175
+ else:
176
+ window = get_window(win_type, win_len, fftbins=True) # **0.5
177
+
178
+ N = fft_len
179
+ fourier_basis = np.fft.rfft(np.eye(N))[:win_len]
180
+ real_kernel = np.real(fourier_basis)
181
+ imag_kernel = np.imag(fourier_basis)
182
+ kernel = np.concatenate([real_kernel, imag_kernel], 1).T
183
+
184
+ if invers:
185
+ kernel = np.linalg.pinv(kernel).T
186
+
187
+ kernel = kernel * window
188
+ kernel = kernel[:, None, :]
189
+ return torch.from_numpy(kernel.astype(np.float32)), torch.from_numpy(window[None, :, None].astype(np.float32))
190
+
modules/losses.py ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch.nn import functional as F
3
+
4
+ import modules.commons as commons
5
+
6
+
7
+ def feature_loss(fmap_r, fmap_g):
8
+ loss = 0
9
+ for dr, dg in zip(fmap_r, fmap_g):
10
+ for rl, gl in zip(dr, dg):
11
+ rl = rl.float().detach()
12
+ gl = gl.float()
13
+ loss += torch.mean(torch.abs(rl - gl))
14
+
15
+ return loss * 2
16
+
17
+
18
+ def discriminator_loss(disc_real_outputs, disc_generated_outputs):
19
+ loss = 0
20
+ r_losses = []
21
+ g_losses = []
22
+ for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
23
+ dr = dr.float()
24
+ dg = dg.float()
25
+ r_loss = torch.mean((1-dr)**2)
26
+ g_loss = torch.mean(dg**2)
27
+ loss += (r_loss + g_loss)
28
+ r_losses.append(r_loss.item())
29
+ g_losses.append(g_loss.item())
30
+
31
+ return loss, r_losses, g_losses
32
+
33
+
34
+ def generator_loss(disc_outputs):
35
+ loss = 0
36
+ gen_losses = []
37
+ for dg in disc_outputs:
38
+ dg = dg.float()
39
+ l = torch.mean((1-dg)**2)
40
+ gen_losses.append(l)
41
+ loss += l
42
+
43
+ return loss, gen_losses
44
+
45
+
46
+ def kl_loss(z_p, logs_q, m_p, logs_p, z_mask):
47
+ """
48
+ z_p, logs_q: [b, h, t_t]
49
+ m_p, logs_p: [b, h, t_t]
50
+ """
51
+ z_p = z_p.float()
52
+ logs_q = logs_q.float()
53
+ m_p = m_p.float()
54
+ logs_p = logs_p.float()
55
+ z_mask = z_mask.float()
56
+ #print(logs_p)
57
+ kl = logs_p - logs_q - 0.5
58
+ kl += 0.5 * ((z_p - m_p)**2) * torch.exp(-2. * logs_p)
59
+ kl = torch.sum(kl * z_mask)
60
+ l = kl / torch.sum(z_mask)
61
+ return l
modules/mel_processing.py ADDED
@@ -0,0 +1,112 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import os
3
+ import random
4
+ import torch
5
+ from torch import nn
6
+ import torch.nn.functional as F
7
+ import torch.utils.data
8
+ import numpy as np
9
+ import librosa
10
+ import librosa.util as librosa_util
11
+ from librosa.util import normalize, pad_center, tiny
12
+ from scipy.signal import get_window
13
+ from scipy.io.wavfile import read
14
+ from librosa.filters import mel as librosa_mel_fn
15
+
16
+ MAX_WAV_VALUE = 32768.0
17
+
18
+
19
+ def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
20
+ """
21
+ PARAMS
22
+ ------
23
+ C: compression factor
24
+ """
25
+ return torch.log(torch.clamp(x, min=clip_val) * C)
26
+
27
+
28
+ def dynamic_range_decompression_torch(x, C=1):
29
+ """
30
+ PARAMS
31
+ ------
32
+ C: compression factor used to compress
33
+ """
34
+ return torch.exp(x) / C
35
+
36
+
37
+ def spectral_normalize_torch(magnitudes):
38
+ output = dynamic_range_compression_torch(magnitudes)
39
+ return output
40
+
41
+
42
+ def spectral_de_normalize_torch(magnitudes):
43
+ output = dynamic_range_decompression_torch(magnitudes)
44
+ return output
45
+
46
+
47
+ mel_basis = {}
48
+ hann_window = {}
49
+
50
+
51
+ def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
52
+ if torch.min(y) < -1.:
53
+ print('min value is ', torch.min(y))
54
+ if torch.max(y) > 1.:
55
+ print('max value is ', torch.max(y))
56
+
57
+ global hann_window
58
+ dtype_device = str(y.dtype) + '_' + str(y.device)
59
+ wnsize_dtype_device = str(win_size) + '_' + dtype_device
60
+ if wnsize_dtype_device not in hann_window:
61
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
62
+
63
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
64
+ y = y.squeeze(1)
65
+
66
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
67
+ center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
68
+
69
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
70
+ return spec
71
+
72
+
73
+ def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
74
+ global mel_basis
75
+ dtype_device = str(spec.dtype) + '_' + str(spec.device)
76
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
77
+ if fmax_dtype_device not in mel_basis:
78
+ mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)
79
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device)
80
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
81
+ spec = spectral_normalize_torch(spec)
82
+ return spec
83
+
84
+
85
+ def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
86
+ if torch.min(y) < -1.:
87
+ print('min value is ', torch.min(y))
88
+ if torch.max(y) > 1.:
89
+ print('max value is ', torch.max(y))
90
+
91
+ global mel_basis, hann_window
92
+ dtype_device = str(y.dtype) + '_' + str(y.device)
93
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
94
+ wnsize_dtype_device = str(win_size) + '_' + dtype_device
95
+ if fmax_dtype_device not in mel_basis:
96
+ mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)
97
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device)
98
+ if wnsize_dtype_device not in hann_window:
99
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
100
+
101
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
102
+ y = y.squeeze(1)
103
+
104
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
105
+ center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
106
+
107
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
108
+
109
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
110
+ spec = spectral_normalize_torch(spec)
111
+
112
+ return spec
modules/modules.py ADDED
@@ -0,0 +1,342 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import math
3
+ import numpy as np
4
+ import scipy
5
+ import torch
6
+ from torch import nn
7
+ from torch.nn import functional as F
8
+
9
+ from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
10
+ from torch.nn.utils import weight_norm, remove_weight_norm
11
+
12
+ import modules.commons as commons
13
+ from modules.commons import init_weights, get_padding
14
+
15
+
16
+ LRELU_SLOPE = 0.1
17
+
18
+
19
+ class LayerNorm(nn.Module):
20
+ def __init__(self, channels, eps=1e-5):
21
+ super().__init__()
22
+ self.channels = channels
23
+ self.eps = eps
24
+
25
+ self.gamma = nn.Parameter(torch.ones(channels))
26
+ self.beta = nn.Parameter(torch.zeros(channels))
27
+
28
+ def forward(self, x):
29
+ x = x.transpose(1, -1)
30
+ x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
31
+ return x.transpose(1, -1)
32
+
33
+
34
+ class ConvReluNorm(nn.Module):
35
+ def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout):
36
+ super().__init__()
37
+ self.in_channels = in_channels
38
+ self.hidden_channels = hidden_channels
39
+ self.out_channels = out_channels
40
+ self.kernel_size = kernel_size
41
+ self.n_layers = n_layers
42
+ self.p_dropout = p_dropout
43
+ assert n_layers > 1, "Number of layers should be larger than 0."
44
+
45
+ self.conv_layers = nn.ModuleList()
46
+ self.norm_layers = nn.ModuleList()
47
+ self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size//2))
48
+ self.norm_layers.append(LayerNorm(hidden_channels))
49
+ self.relu_drop = nn.Sequential(
50
+ nn.ReLU(),
51
+ nn.Dropout(p_dropout))
52
+ for _ in range(n_layers-1):
53
+ self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size//2))
54
+ self.norm_layers.append(LayerNorm(hidden_channels))
55
+ self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
56
+ self.proj.weight.data.zero_()
57
+ self.proj.bias.data.zero_()
58
+
59
+ def forward(self, x, x_mask):
60
+ x_org = x
61
+ for i in range(self.n_layers):
62
+ x = self.conv_layers[i](x * x_mask)
63
+ x = self.norm_layers[i](x)
64
+ x = self.relu_drop(x)
65
+ x = x_org + self.proj(x)
66
+ return x * x_mask
67
+
68
+
69
+ class DDSConv(nn.Module):
70
+ """
71
+ Dialted and Depth-Separable Convolution
72
+ """
73
+ def __init__(self, channels, kernel_size, n_layers, p_dropout=0.):
74
+ super().__init__()
75
+ self.channels = channels
76
+ self.kernel_size = kernel_size
77
+ self.n_layers = n_layers
78
+ self.p_dropout = p_dropout
79
+
80
+ self.drop = nn.Dropout(p_dropout)
81
+ self.convs_sep = nn.ModuleList()
82
+ self.convs_1x1 = nn.ModuleList()
83
+ self.norms_1 = nn.ModuleList()
84
+ self.norms_2 = nn.ModuleList()
85
+ for i in range(n_layers):
86
+ dilation = kernel_size ** i
87
+ padding = (kernel_size * dilation - dilation) // 2
88
+ self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size,
89
+ groups=channels, dilation=dilation, padding=padding
90
+ ))
91
+ self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
92
+ self.norms_1.append(LayerNorm(channels))
93
+ self.norms_2.append(LayerNorm(channels))
94
+
95
+ def forward(self, x, x_mask, g=None):
96
+ if g is not None:
97
+ x = x + g
98
+ for i in range(self.n_layers):
99
+ y = self.convs_sep[i](x * x_mask)
100
+ y = self.norms_1[i](y)
101
+ y = F.gelu(y)
102
+ y = self.convs_1x1[i](y)
103
+ y = self.norms_2[i](y)
104
+ y = F.gelu(y)
105
+ y = self.drop(y)
106
+ x = x + y
107
+ return x * x_mask
108
+
109
+
110
+ class WN(torch.nn.Module):
111
+ def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
112
+ super(WN, self).__init__()
113
+ assert(kernel_size % 2 == 1)
114
+ self.hidden_channels =hidden_channels
115
+ self.kernel_size = kernel_size,
116
+ self.dilation_rate = dilation_rate
117
+ self.n_layers = n_layers
118
+ self.gin_channels = gin_channels
119
+ self.p_dropout = p_dropout
120
+
121
+ self.in_layers = torch.nn.ModuleList()
122
+ self.res_skip_layers = torch.nn.ModuleList()
123
+ self.drop = nn.Dropout(p_dropout)
124
+
125
+ if gin_channels != 0:
126
+ cond_layer = torch.nn.Conv1d(gin_channels, 2*hidden_channels*n_layers, 1)
127
+ self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name='weight')
128
+
129
+ for i in range(n_layers):
130
+ dilation = dilation_rate ** i
131
+ padding = int((kernel_size * dilation - dilation) / 2)
132
+ in_layer = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, kernel_size,
133
+ dilation=dilation, padding=padding)
134
+ in_layer = torch.nn.utils.weight_norm(in_layer, name='weight')
135
+ self.in_layers.append(in_layer)
136
+
137
+ # last one is not necessary
138
+ if i < n_layers - 1:
139
+ res_skip_channels = 2 * hidden_channels
140
+ else:
141
+ res_skip_channels = hidden_channels
142
+
143
+ res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
144
+ res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name='weight')
145
+ self.res_skip_layers.append(res_skip_layer)
146
+
147
+ def forward(self, x, x_mask, g=None, **kwargs):
148
+ output = torch.zeros_like(x)
149
+ n_channels_tensor = torch.IntTensor([self.hidden_channels])
150
+
151
+ if g is not None:
152
+ g = self.cond_layer(g)
153
+
154
+ for i in range(self.n_layers):
155
+ x_in = self.in_layers[i](x)
156
+ if g is not None:
157
+ cond_offset = i * 2 * self.hidden_channels
158
+ g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:]
159
+ else:
160
+ g_l = torch.zeros_like(x_in)
161
+
162
+ acts = commons.fused_add_tanh_sigmoid_multiply(
163
+ x_in,
164
+ g_l,
165
+ n_channels_tensor)
166
+ acts = self.drop(acts)
167
+
168
+ res_skip_acts = self.res_skip_layers[i](acts)
169
+ if i < self.n_layers - 1:
170
+ res_acts = res_skip_acts[:,:self.hidden_channels,:]
171
+ x = (x + res_acts) * x_mask
172
+ output = output + res_skip_acts[:,self.hidden_channels:,:]
173
+ else:
174
+ output = output + res_skip_acts
175
+ return output * x_mask
176
+
177
+ def remove_weight_norm(self):
178
+ if self.gin_channels != 0:
179
+ torch.nn.utils.remove_weight_norm(self.cond_layer)
180
+ for l in self.in_layers:
181
+ torch.nn.utils.remove_weight_norm(l)
182
+ for l in self.res_skip_layers:
183
+ torch.nn.utils.remove_weight_norm(l)
184
+
185
+
186
+ class ResBlock1(torch.nn.Module):
187
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
188
+ super(ResBlock1, self).__init__()
189
+ self.convs1 = nn.ModuleList([
190
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
191
+ padding=get_padding(kernel_size, dilation[0]))),
192
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
193
+ padding=get_padding(kernel_size, dilation[1]))),
194
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
195
+ padding=get_padding(kernel_size, dilation[2])))
196
+ ])
197
+ self.convs1.apply(init_weights)
198
+
199
+ self.convs2 = nn.ModuleList([
200
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
201
+ padding=get_padding(kernel_size, 1))),
202
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
203
+ padding=get_padding(kernel_size, 1))),
204
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
205
+ padding=get_padding(kernel_size, 1)))
206
+ ])
207
+ self.convs2.apply(init_weights)
208
+
209
+ def forward(self, x, x_mask=None):
210
+ for c1, c2 in zip(self.convs1, self.convs2):
211
+ xt = F.leaky_relu(x, LRELU_SLOPE)
212
+ if x_mask is not None:
213
+ xt = xt * x_mask
214
+ xt = c1(xt)
215
+ xt = F.leaky_relu(xt, LRELU_SLOPE)
216
+ if x_mask is not None:
217
+ xt = xt * x_mask
218
+ xt = c2(xt)
219
+ x = xt + x
220
+ if x_mask is not None:
221
+ x = x * x_mask
222
+ return x
223
+
224
+ def remove_weight_norm(self):
225
+ for l in self.convs1:
226
+ remove_weight_norm(l)
227
+ for l in self.convs2:
228
+ remove_weight_norm(l)
229
+
230
+
231
+ class ResBlock2(torch.nn.Module):
232
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
233
+ super(ResBlock2, self).__init__()
234
+ self.convs = nn.ModuleList([
235
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
236
+ padding=get_padding(kernel_size, dilation[0]))),
237
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
238
+ padding=get_padding(kernel_size, dilation[1])))
239
+ ])
240
+ self.convs.apply(init_weights)
241
+
242
+ def forward(self, x, x_mask=None):
243
+ for c in self.convs:
244
+ xt = F.leaky_relu(x, LRELU_SLOPE)
245
+ if x_mask is not None:
246
+ xt = xt * x_mask
247
+ xt = c(xt)
248
+ x = xt + x
249
+ if x_mask is not None:
250
+ x = x * x_mask
251
+ return x
252
+
253
+ def remove_weight_norm(self):
254
+ for l in self.convs:
255
+ remove_weight_norm(l)
256
+
257
+
258
+ class Log(nn.Module):
259
+ def forward(self, x, x_mask, reverse=False, **kwargs):
260
+ if not reverse:
261
+ y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
262
+ logdet = torch.sum(-y, [1, 2])
263
+ return y, logdet
264
+ else:
265
+ x = torch.exp(x) * x_mask
266
+ return x
267
+
268
+
269
+ class Flip(nn.Module):
270
+ def forward(self, x, *args, reverse=False, **kwargs):
271
+ x = torch.flip(x, [1])
272
+ if not reverse:
273
+ logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
274
+ return x, logdet
275
+ else:
276
+ return x
277
+
278
+
279
+ class ElementwiseAffine(nn.Module):
280
+ def __init__(self, channels):
281
+ super().__init__()
282
+ self.channels = channels
283
+ self.m = nn.Parameter(torch.zeros(channels,1))
284
+ self.logs = nn.Parameter(torch.zeros(channels,1))
285
+
286
+ def forward(self, x, x_mask, reverse=False, **kwargs):
287
+ if not reverse:
288
+ y = self.m + torch.exp(self.logs) * x
289
+ y = y * x_mask
290
+ logdet = torch.sum(self.logs * x_mask, [1,2])
291
+ return y, logdet
292
+ else:
293
+ x = (x - self.m) * torch.exp(-self.logs) * x_mask
294
+ return x
295
+
296
+
297
+ class ResidualCouplingLayer(nn.Module):
298
+ def __init__(self,
299
+ channels,
300
+ hidden_channels,
301
+ kernel_size,
302
+ dilation_rate,
303
+ n_layers,
304
+ p_dropout=0,
305
+ gin_channels=0,
306
+ mean_only=False):
307
+ assert channels % 2 == 0, "channels should be divisible by 2"
308
+ super().__init__()
309
+ self.channels = channels
310
+ self.hidden_channels = hidden_channels
311
+ self.kernel_size = kernel_size
312
+ self.dilation_rate = dilation_rate
313
+ self.n_layers = n_layers
314
+ self.half_channels = channels // 2
315
+ self.mean_only = mean_only
316
+
317
+ self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
318
+ self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
319
+ self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
320
+ self.post.weight.data.zero_()
321
+ self.post.bias.data.zero_()
322
+
323
+ def forward(self, x, x_mask, g=None, reverse=False):
324
+ x0, x1 = torch.split(x, [self.half_channels]*2, 1)
325
+ h = self.pre(x0) * x_mask
326
+ h = self.enc(h, x_mask, g=g)
327
+ stats = self.post(h) * x_mask
328
+ if not self.mean_only:
329
+ m, logs = torch.split(stats, [self.half_channels]*2, 1)
330
+ else:
331
+ m = stats
332
+ logs = torch.zeros_like(m)
333
+
334
+ if not reverse:
335
+ x1 = m + x1 * torch.exp(logs) * x_mask
336
+ x = torch.cat([x0, x1], 1)
337
+ logdet = torch.sum(logs, [1,2])
338
+ return x, logdet
339
+ else:
340
+ x1 = (x1 - m) * torch.exp(-logs) * x_mask
341
+ x = torch.cat([x0, x1], 1)
342
+ return x
onnx/model_onnx.py ADDED
@@ -0,0 +1,328 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import math
3
+ import torch
4
+ from torch import nn
5
+ from torch.nn import functional as F
6
+
7
+ import modules.attentions as attentions
8
+ import modules.commons as commons
9
+ import modules.modules as modules
10
+
11
+ from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
12
+ from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
13
+ from modules.commons import init_weights, get_padding
14
+ from vdecoder.hifigan.models import Generator
15
+ from utils import f0_to_coarse
16
+
17
+ class ResidualCouplingBlock(nn.Module):
18
+ def __init__(self,
19
+ channels,
20
+ hidden_channels,
21
+ kernel_size,
22
+ dilation_rate,
23
+ n_layers,
24
+ n_flows=4,
25
+ gin_channels=0):
26
+ super().__init__()
27
+ self.channels = channels
28
+ self.hidden_channels = hidden_channels
29
+ self.kernel_size = kernel_size
30
+ self.dilation_rate = dilation_rate
31
+ self.n_layers = n_layers
32
+ self.n_flows = n_flows
33
+ self.gin_channels = gin_channels
34
+
35
+ self.flows = nn.ModuleList()
36
+ for i in range(n_flows):
37
+ self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
38
+ self.flows.append(modules.Flip())
39
+
40
+ def forward(self, x, x_mask, g=None, reverse=False):
41
+ if not reverse:
42
+ for flow in self.flows:
43
+ x, _ = flow(x, x_mask, g=g, reverse=reverse)
44
+ else:
45
+ for flow in reversed(self.flows):
46
+ x = flow(x, x_mask, g=g, reverse=reverse)
47
+ return x
48
+
49
+
50
+ class Encoder(nn.Module):
51
+ def __init__(self,
52
+ in_channels,
53
+ out_channels,
54
+ hidden_channels,
55
+ kernel_size,
56
+ dilation_rate,
57
+ n_layers,
58
+ gin_channels=0):
59
+ super().__init__()
60
+ self.in_channels = in_channels
61
+ self.out_channels = out_channels
62
+ self.hidden_channels = hidden_channels
63
+ self.kernel_size = kernel_size
64
+ self.dilation_rate = dilation_rate
65
+ self.n_layers = n_layers
66
+ self.gin_channels = gin_channels
67
+
68
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
69
+ self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
70
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
71
+
72
+ def forward(self, x, x_lengths, g=None):
73
+ # print(x.shape,x_lengths.shape)
74
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
75
+ x = self.pre(x) * x_mask
76
+ x = self.enc(x, x_mask, g=g)
77
+ stats = self.proj(x) * x_mask
78
+ m, logs = torch.split(stats, self.out_channels, dim=1)
79
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
80
+ return z, m, logs, x_mask
81
+
82
+
83
+ class TextEncoder(nn.Module):
84
+ def __init__(self,
85
+ in_channels,
86
+ out_channels,
87
+ hidden_channels,
88
+ kernel_size,
89
+ dilation_rate,
90
+ n_layers,
91
+ gin_channels=0,
92
+ filter_channels=None,
93
+ n_heads=None,
94
+ p_dropout=None):
95
+ super().__init__()
96
+ self.in_channels = in_channels
97
+ self.out_channels = out_channels
98
+ self.hidden_channels = hidden_channels
99
+ self.kernel_size = kernel_size
100
+ self.dilation_rate = dilation_rate
101
+ self.n_layers = n_layers
102
+ self.gin_channels = gin_channels
103
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
104
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
105
+ self.f0_emb = nn.Embedding(256, hidden_channels)
106
+
107
+ self.enc_ = attentions.Encoder(
108
+ hidden_channels,
109
+ filter_channels,
110
+ n_heads,
111
+ n_layers,
112
+ kernel_size,
113
+ p_dropout)
114
+
115
+ def forward(self, x, x_lengths, f0=None):
116
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
117
+ x = self.pre(x) * x_mask
118
+ x = x + self.f0_emb(f0.long()).transpose(1,2)
119
+ x = self.enc_(x * x_mask, x_mask)
120
+ stats = self.proj(x) * x_mask
121
+ m, logs = torch.split(stats, self.out_channels, dim=1)
122
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
123
+
124
+ return z, m, logs, x_mask
125
+
126
+
127
+
128
+ class DiscriminatorP(torch.nn.Module):
129
+ def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
130
+ super(DiscriminatorP, self).__init__()
131
+ self.period = period
132
+ self.use_spectral_norm = use_spectral_norm
133
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
134
+ self.convs = nn.ModuleList([
135
+ norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
136
+ norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
137
+ norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
138
+ norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
139
+ norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
140
+ ])
141
+ self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
142
+
143
+ def forward(self, x):
144
+ fmap = []
145
+
146
+ # 1d to 2d
147
+ b, c, t = x.shape
148
+ if t % self.period != 0: # pad first
149
+ n_pad = self.period - (t % self.period)
150
+ x = F.pad(x, (0, n_pad), "reflect")
151
+ t = t + n_pad
152
+ x = x.view(b, c, t // self.period, self.period)
153
+
154
+ for l in self.convs:
155
+ x = l(x)
156
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
157
+ fmap.append(x)
158
+ x = self.conv_post(x)
159
+ fmap.append(x)
160
+ x = torch.flatten(x, 1, -1)
161
+
162
+ return x, fmap
163
+
164
+
165
+ class DiscriminatorS(torch.nn.Module):
166
+ def __init__(self, use_spectral_norm=False):
167
+ super(DiscriminatorS, self).__init__()
168
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
169
+ self.convs = nn.ModuleList([
170
+ norm_f(Conv1d(1, 16, 15, 1, padding=7)),
171
+ norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
172
+ norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
173
+ norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
174
+ norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
175
+ norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
176
+ ])
177
+ self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
178
+
179
+ def forward(self, x):
180
+ fmap = []
181
+
182
+ for l in self.convs:
183
+ x = l(x)
184
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
185
+ fmap.append(x)
186
+ x = self.conv_post(x)
187
+ fmap.append(x)
188
+ x = torch.flatten(x, 1, -1)
189
+
190
+ return x, fmap
191
+
192
+
193
+ class MultiPeriodDiscriminator(torch.nn.Module):
194
+ def __init__(self, use_spectral_norm=False):
195
+ super(MultiPeriodDiscriminator, self).__init__()
196
+ periods = [2,3,5,7,11]
197
+
198
+ discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
199
+ discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
200
+ self.discriminators = nn.ModuleList(discs)
201
+
202
+ def forward(self, y, y_hat):
203
+ y_d_rs = []
204
+ y_d_gs = []
205
+ fmap_rs = []
206
+ fmap_gs = []
207
+ for i, d in enumerate(self.discriminators):
208
+ y_d_r, fmap_r = d(y)
209
+ y_d_g, fmap_g = d(y_hat)
210
+ y_d_rs.append(y_d_r)
211
+ y_d_gs.append(y_d_g)
212
+ fmap_rs.append(fmap_r)
213
+ fmap_gs.append(fmap_g)
214
+
215
+ return y_d_rs, y_d_gs, fmap_rs, fmap_gs
216
+
217
+
218
+ class SpeakerEncoder(torch.nn.Module):
219
+ def __init__(self, mel_n_channels=80, model_num_layers=3, model_hidden_size=256, model_embedding_size=256):
220
+ super(SpeakerEncoder, self).__init__()
221
+ self.lstm = nn.LSTM(mel_n_channels, model_hidden_size, model_num_layers, batch_first=True)
222
+ self.linear = nn.Linear(model_hidden_size, model_embedding_size)
223
+ self.relu = nn.ReLU()
224
+
225
+ def forward(self, mels):
226
+ self.lstm.flatten_parameters()
227
+ _, (hidden, _) = self.lstm(mels)
228
+ embeds_raw = self.relu(self.linear(hidden[-1]))
229
+ return embeds_raw / torch.norm(embeds_raw, dim=1, keepdim=True)
230
+
231
+ def compute_partial_slices(self, total_frames, partial_frames, partial_hop):
232
+ mel_slices = []
233
+ for i in range(0, total_frames-partial_frames, partial_hop):
234
+ mel_range = torch.arange(i, i+partial_frames)
235
+ mel_slices.append(mel_range)
236
+
237
+ return mel_slices
238
+
239
+ def embed_utterance(self, mel, partial_frames=128, partial_hop=64):
240
+ mel_len = mel.size(1)
241
+ last_mel = mel[:,-partial_frames:]
242
+
243
+ if mel_len > partial_frames:
244
+ mel_slices = self.compute_partial_slices(mel_len, partial_frames, partial_hop)
245
+ mels = list(mel[:,s] for s in mel_slices)
246
+ mels.append(last_mel)
247
+ mels = torch.stack(tuple(mels), 0).squeeze(1)
248
+
249
+ with torch.no_grad():
250
+ partial_embeds = self(mels)
251
+ embed = torch.mean(partial_embeds, axis=0).unsqueeze(0)
252
+ #embed = embed / torch.linalg.norm(embed, 2)
253
+ else:
254
+ with torch.no_grad():
255
+ embed = self(last_mel)
256
+
257
+ return embed
258
+
259
+
260
+ class SynthesizerTrn(nn.Module):
261
+ """
262
+ Synthesizer for Training
263
+ """
264
+
265
+ def __init__(self,
266
+ spec_channels,
267
+ segment_size,
268
+ inter_channels,
269
+ hidden_channels,
270
+ filter_channels,
271
+ n_heads,
272
+ n_layers,
273
+ kernel_size,
274
+ p_dropout,
275
+ resblock,
276
+ resblock_kernel_sizes,
277
+ resblock_dilation_sizes,
278
+ upsample_rates,
279
+ upsample_initial_channel,
280
+ upsample_kernel_sizes,
281
+ gin_channels,
282
+ ssl_dim,
283
+ n_speakers,
284
+ **kwargs):
285
+
286
+ super().__init__()
287
+ self.spec_channels = spec_channels
288
+ self.inter_channels = inter_channels
289
+ self.hidden_channels = hidden_channels
290
+ self.filter_channels = filter_channels
291
+ self.n_heads = n_heads
292
+ self.n_layers = n_layers
293
+ self.kernel_size = kernel_size
294
+ self.p_dropout = p_dropout
295
+ self.resblock = resblock
296
+ self.resblock_kernel_sizes = resblock_kernel_sizes
297
+ self.resblock_dilation_sizes = resblock_dilation_sizes
298
+ self.upsample_rates = upsample_rates
299
+ self.upsample_initial_channel = upsample_initial_channel
300
+ self.upsample_kernel_sizes = upsample_kernel_sizes
301
+ self.segment_size = segment_size
302
+ self.gin_channels = gin_channels
303
+ self.ssl_dim = ssl_dim
304
+ self.emb_g = nn.Embedding(n_speakers, gin_channels)
305
+
306
+ self.enc_p_ = TextEncoder(ssl_dim, inter_channels, hidden_channels, 5, 1, 16,0, filter_channels, n_heads, p_dropout)
307
+ hps = {
308
+ "sampling_rate": 32000,
309
+ "inter_channels": 192,
310
+ "resblock": "1",
311
+ "resblock_kernel_sizes": [3, 7, 11],
312
+ "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
313
+ "upsample_rates": [10, 8, 2, 2],
314
+ "upsample_initial_channel": 512,
315
+ "upsample_kernel_sizes": [16, 16, 4, 4],
316
+ "gin_channels": 256,
317
+ }
318
+ self.dec = Generator(h=hps)
319
+ self.enc_q = Encoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
320
+ self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
321
+
322
+ def forward(self, c, c_lengths, f0, g=None):
323
+ g = self.emb_g(g.unsqueeze(0)).transpose(1,2)
324
+ z_p, m_p, logs_p, c_mask = self.enc_p_(c.transpose(1,2), c_lengths, f0=f0_to_coarse(f0))
325
+ z = self.flow(z_p, c_mask, g=g, reverse=True)
326
+ o = self.dec(z * c_mask, g=g, f0=f0.float())
327
+ return o
328
+
onnx/model_onnx_48k.py ADDED
@@ -0,0 +1,328 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import math
3
+ import torch
4
+ from torch import nn
5
+ from torch.nn import functional as F
6
+
7
+ import modules.attentions as attentions
8
+ import modules.commons as commons
9
+ import modules.modules as modules
10
+
11
+ from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
12
+ from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
13
+ from modules.commons import init_weights, get_padding
14
+ from vdecoder.hifigan.models import Generator
15
+ from utils import f0_to_coarse
16
+
17
+ class ResidualCouplingBlock(nn.Module):
18
+ def __init__(self,
19
+ channels,
20
+ hidden_channels,
21
+ kernel_size,
22
+ dilation_rate,
23
+ n_layers,
24
+ n_flows=4,
25
+ gin_channels=0):
26
+ super().__init__()
27
+ self.channels = channels
28
+ self.hidden_channels = hidden_channels
29
+ self.kernel_size = kernel_size
30
+ self.dilation_rate = dilation_rate
31
+ self.n_layers = n_layers
32
+ self.n_flows = n_flows
33
+ self.gin_channels = gin_channels
34
+
35
+ self.flows = nn.ModuleList()
36
+ for i in range(n_flows):
37
+ self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
38
+ self.flows.append(modules.Flip())
39
+
40
+ def forward(self, x, x_mask, g=None, reverse=False):
41
+ if not reverse:
42
+ for flow in self.flows:
43
+ x, _ = flow(x, x_mask, g=g, reverse=reverse)
44
+ else:
45
+ for flow in reversed(self.flows):
46
+ x = flow(x, x_mask, g=g, reverse=reverse)
47
+ return x
48
+
49
+
50
+ class Encoder(nn.Module):
51
+ def __init__(self,
52
+ in_channels,
53
+ out_channels,
54
+ hidden_channels,
55
+ kernel_size,
56
+ dilation_rate,
57
+ n_layers,
58
+ gin_channels=0):
59
+ super().__init__()
60
+ self.in_channels = in_channels
61
+ self.out_channels = out_channels
62
+ self.hidden_channels = hidden_channels
63
+ self.kernel_size = kernel_size
64
+ self.dilation_rate = dilation_rate
65
+ self.n_layers = n_layers
66
+ self.gin_channels = gin_channels
67
+
68
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
69
+ self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
70
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
71
+
72
+ def forward(self, x, x_lengths, g=None):
73
+ # print(x.shape,x_lengths.shape)
74
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
75
+ x = self.pre(x) * x_mask
76
+ x = self.enc(x, x_mask, g=g)
77
+ stats = self.proj(x) * x_mask
78
+ m, logs = torch.split(stats, self.out_channels, dim=1)
79
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
80
+ return z, m, logs, x_mask
81
+
82
+
83
+ class TextEncoder(nn.Module):
84
+ def __init__(self,
85
+ in_channels,
86
+ out_channels,
87
+ hidden_channels,
88
+ kernel_size,
89
+ dilation_rate,
90
+ n_layers,
91
+ gin_channels=0,
92
+ filter_channels=None,
93
+ n_heads=None,
94
+ p_dropout=None):
95
+ super().__init__()
96
+ self.in_channels = in_channels
97
+ self.out_channels = out_channels
98
+ self.hidden_channels = hidden_channels
99
+ self.kernel_size = kernel_size
100
+ self.dilation_rate = dilation_rate
101
+ self.n_layers = n_layers
102
+ self.gin_channels = gin_channels
103
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
104
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
105
+ self.f0_emb = nn.Embedding(256, hidden_channels)
106
+
107
+ self.enc_ = attentions.Encoder(
108
+ hidden_channels,
109
+ filter_channels,
110
+ n_heads,
111
+ n_layers,
112
+ kernel_size,
113
+ p_dropout)
114
+
115
+ def forward(self, x, x_lengths, f0=None):
116
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
117
+ x = self.pre(x) * x_mask
118
+ x = x + self.f0_emb(f0.long()).transpose(1,2)
119
+ x = self.enc_(x * x_mask, x_mask)
120
+ stats = self.proj(x) * x_mask
121
+ m, logs = torch.split(stats, self.out_channels, dim=1)
122
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
123
+
124
+ return z, m, logs, x_mask
125
+
126
+
127
+
128
+ class DiscriminatorP(torch.nn.Module):
129
+ def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
130
+ super(DiscriminatorP, self).__init__()
131
+ self.period = period
132
+ self.use_spectral_norm = use_spectral_norm
133
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
134
+ self.convs = nn.ModuleList([
135
+ norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
136
+ norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
137
+ norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
138
+ norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
139
+ norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
140
+ ])
141
+ self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
142
+
143
+ def forward(self, x):
144
+ fmap = []
145
+
146
+ # 1d to 2d
147
+ b, c, t = x.shape
148
+ if t % self.period != 0: # pad first
149
+ n_pad = self.period - (t % self.period)
150
+ x = F.pad(x, (0, n_pad), "reflect")
151
+ t = t + n_pad
152
+ x = x.view(b, c, t // self.period, self.period)
153
+
154
+ for l in self.convs:
155
+ x = l(x)
156
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
157
+ fmap.append(x)
158
+ x = self.conv_post(x)
159
+ fmap.append(x)
160
+ x = torch.flatten(x, 1, -1)
161
+
162
+ return x, fmap
163
+
164
+
165
+ class DiscriminatorS(torch.nn.Module):
166
+ def __init__(self, use_spectral_norm=False):
167
+ super(DiscriminatorS, self).__init__()
168
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
169
+ self.convs = nn.ModuleList([
170
+ norm_f(Conv1d(1, 16, 15, 1, padding=7)),
171
+ norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
172
+ norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
173
+ norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
174
+ norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
175
+ norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
176
+ ])
177
+ self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
178
+
179
+ def forward(self, x):
180
+ fmap = []
181
+
182
+ for l in self.convs:
183
+ x = l(x)
184
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
185
+ fmap.append(x)
186
+ x = self.conv_post(x)
187
+ fmap.append(x)
188
+ x = torch.flatten(x, 1, -1)
189
+
190
+ return x, fmap
191
+
192
+
193
+ class MultiPeriodDiscriminator(torch.nn.Module):
194
+ def __init__(self, use_spectral_norm=False):
195
+ super(MultiPeriodDiscriminator, self).__init__()
196
+ periods = [2,3,5,7,11]
197
+
198
+ discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
199
+ discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
200
+ self.discriminators = nn.ModuleList(discs)
201
+
202
+ def forward(self, y, y_hat):
203
+ y_d_rs = []
204
+ y_d_gs = []
205
+ fmap_rs = []
206
+ fmap_gs = []
207
+ for i, d in enumerate(self.discriminators):
208
+ y_d_r, fmap_r = d(y)
209
+ y_d_g, fmap_g = d(y_hat)
210
+ y_d_rs.append(y_d_r)
211
+ y_d_gs.append(y_d_g)
212
+ fmap_rs.append(fmap_r)
213
+ fmap_gs.append(fmap_g)
214
+
215
+ return y_d_rs, y_d_gs, fmap_rs, fmap_gs
216
+
217
+
218
+ class SpeakerEncoder(torch.nn.Module):
219
+ def __init__(self, mel_n_channels=80, model_num_layers=3, model_hidden_size=256, model_embedding_size=256):
220
+ super(SpeakerEncoder, self).__init__()
221
+ self.lstm = nn.LSTM(mel_n_channels, model_hidden_size, model_num_layers, batch_first=True)
222
+ self.linear = nn.Linear(model_hidden_size, model_embedding_size)
223
+ self.relu = nn.ReLU()
224
+
225
+ def forward(self, mels):
226
+ self.lstm.flatten_parameters()
227
+ _, (hidden, _) = self.lstm(mels)
228
+ embeds_raw = self.relu(self.linear(hidden[-1]))
229
+ return embeds_raw / torch.norm(embeds_raw, dim=1, keepdim=True)
230
+
231
+ def compute_partial_slices(self, total_frames, partial_frames, partial_hop):
232
+ mel_slices = []
233
+ for i in range(0, total_frames-partial_frames, partial_hop):
234
+ mel_range = torch.arange(i, i+partial_frames)
235
+ mel_slices.append(mel_range)
236
+
237
+ return mel_slices
238
+
239
+ def embed_utterance(self, mel, partial_frames=128, partial_hop=64):
240
+ mel_len = mel.size(1)
241
+ last_mel = mel[:,-partial_frames:]
242
+
243
+ if mel_len > partial_frames:
244
+ mel_slices = self.compute_partial_slices(mel_len, partial_frames, partial_hop)
245
+ mels = list(mel[:,s] for s in mel_slices)
246
+ mels.append(last_mel)
247
+ mels = torch.stack(tuple(mels), 0).squeeze(1)
248
+
249
+ with torch.no_grad():
250
+ partial_embeds = self(mels)
251
+ embed = torch.mean(partial_embeds, axis=0).unsqueeze(0)
252
+ #embed = embed / torch.linalg.norm(embed, 2)
253
+ else:
254
+ with torch.no_grad():
255
+ embed = self(last_mel)
256
+
257
+ return embed
258
+
259
+
260
+ class SynthesizerTrn(nn.Module):
261
+ """
262
+ Synthesizer for Training
263
+ """
264
+
265
+ def __init__(self,
266
+ spec_channels,
267
+ segment_size,
268
+ inter_channels,
269
+ hidden_channels,
270
+ filter_channels,
271
+ n_heads,
272
+ n_layers,
273
+ kernel_size,
274
+ p_dropout,
275
+ resblock,
276
+ resblock_kernel_sizes,
277
+ resblock_dilation_sizes,
278
+ upsample_rates,
279
+ upsample_initial_channel,
280
+ upsample_kernel_sizes,
281
+ gin_channels,
282
+ ssl_dim,
283
+ n_speakers,
284
+ **kwargs):
285
+
286
+ super().__init__()
287
+ self.spec_channels = spec_channels
288
+ self.inter_channels = inter_channels
289
+ self.hidden_channels = hidden_channels
290
+ self.filter_channels = filter_channels
291
+ self.n_heads = n_heads
292
+ self.n_layers = n_layers
293
+ self.kernel_size = kernel_size
294
+ self.p_dropout = p_dropout
295
+ self.resblock = resblock
296
+ self.resblock_kernel_sizes = resblock_kernel_sizes
297
+ self.resblock_dilation_sizes = resblock_dilation_sizes
298
+ self.upsample_rates = upsample_rates
299
+ self.upsample_initial_channel = upsample_initial_channel
300
+ self.upsample_kernel_sizes = upsample_kernel_sizes
301
+ self.segment_size = segment_size
302
+ self.gin_channels = gin_channels
303
+ self.ssl_dim = ssl_dim
304
+ self.emb_g = nn.Embedding(n_speakers, gin_channels)
305
+
306
+ self.enc_p_ = TextEncoder(ssl_dim, inter_channels, hidden_channels, 5, 1, 16,0, filter_channels, n_heads, p_dropout)
307
+ hps = {
308
+ "sampling_rate": 48000,
309
+ "inter_channels": 192,
310
+ "resblock": "1",
311
+ "resblock_kernel_sizes": [3, 7, 11],
312
+ "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
313
+ "upsample_rates": [10, 8, 2, 2],
314
+ "upsample_initial_channel": 512,
315
+ "upsample_kernel_sizes": [16, 16, 4, 4],
316
+ "gin_channels": 256,
317
+ }
318
+ self.dec = Generator(h=hps)
319
+ self.enc_q = Encoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
320
+ self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
321
+
322
+ def forward(self, c, c_lengths, f0, g=None):
323
+ g = self.emb_g(g.unsqueeze(0)).transpose(1,2)
324
+ z_p, m_p, logs_p, c_mask = self.enc_p_(c.transpose(1,2), c_lengths, f0=f0_to_coarse(f0))
325
+ z = self.flow(z_p, c_mask, g=g, reverse=True)
326
+ o = self.dec(z * c_mask, g=g, f0=f0.float())
327
+ return o
328
+
onnx/onnx_export.py ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import time
3
+ import numpy as np
4
+ import onnx
5
+ from onnxsim import simplify
6
+ import onnxruntime as ort
7
+ import onnxoptimizer
8
+ import torch
9
+ from model_onnx import SynthesizerTrn
10
+ import utils
11
+ from hubert import hubert_model_onnx
12
+
13
+ def main(HubertExport,NetExport):
14
+
15
+ path = "NyaruTaffy"
16
+
17
+ if(HubertExport):
18
+ device = torch.device("cuda")
19
+ hubert_soft = utils.get_hubert_model()
20
+ test_input = torch.rand(1, 1, 16000)
21
+ input_names = ["source"]
22
+ output_names = ["embed"]
23
+ torch.onnx.export(hubert_soft.to(device),
24
+ test_input.to(device),
25
+ "hubert3.0.onnx",
26
+ dynamic_axes={
27
+ "source": {
28
+ 2: "sample_length"
29
+ }
30
+ },
31
+ verbose=False,
32
+ opset_version=13,
33
+ input_names=input_names,
34
+ output_names=output_names)
35
+ if(NetExport):
36
+ device = torch.device("cuda")
37
+ hps = utils.get_hparams_from_file(f"checkpoints/{path}/config.json")
38
+ SVCVITS = SynthesizerTrn(
39
+ hps.data.filter_length // 2 + 1,
40
+ hps.train.segment_size // hps.data.hop_length,
41
+ **hps.model)
42
+ _ = utils.load_checkpoint(f"checkpoints/{path}/model.pth", SVCVITS, None)
43
+ _ = SVCVITS.eval().to(device)
44
+ for i in SVCVITS.parameters():
45
+ i.requires_grad = False
46
+ test_hidden_unit = torch.rand(1, 50, 256)
47
+ test_lengths = torch.LongTensor([50])
48
+ test_pitch = torch.rand(1, 50)
49
+ test_sid = torch.LongTensor([0])
50
+ input_names = ["hidden_unit", "lengths", "pitch", "sid"]
51
+ output_names = ["audio", ]
52
+ SVCVITS.eval()
53
+ torch.onnx.export(SVCVITS,
54
+ (
55
+ test_hidden_unit.to(device),
56
+ test_lengths.to(device),
57
+ test_pitch.to(device),
58
+ test_sid.to(device)
59
+ ),
60
+ f"checkpoints/{path}/model.onnx",
61
+ dynamic_axes={
62
+ "hidden_unit": [0, 1],
63
+ "pitch": [1]
64
+ },
65
+ do_constant_folding=False,
66
+ opset_version=16,
67
+ verbose=False,
68
+ input_names=input_names,
69
+ output_names=output_names)
70
+
71
+
72
+ if __name__ == '__main__':
73
+ main(False,True)
onnx/onnx_export_48k.py ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import time
3
+ import numpy as np
4
+ import onnx
5
+ from onnxsim import simplify
6
+ import onnxruntime as ort
7
+ import onnxoptimizer
8
+ import torch
9
+ from model_onnx_48k import SynthesizerTrn
10
+ import utils
11
+ from hubert import hubert_model_onnx
12
+
13
+ def main(HubertExport,NetExport):
14
+
15
+ path = "NyaruTaffy"
16
+
17
+ if(HubertExport):
18
+ device = torch.device("cuda")
19
+ hubert_soft = hubert_model_onnx.hubert_soft("hubert/model.pt")
20
+ test_input = torch.rand(1, 1, 16000)
21
+ input_names = ["source"]
22
+ output_names = ["embed"]
23
+ torch.onnx.export(hubert_soft.to(device),
24
+ test_input.to(device),
25
+ "hubert3.0.onnx",
26
+ dynamic_axes={
27
+ "source": {
28
+ 2: "sample_length"
29
+ }
30
+ },
31
+ verbose=False,
32
+ opset_version=13,
33
+ input_names=input_names,
34
+ output_names=output_names)
35
+ if(NetExport):
36
+ device = torch.device("cuda")
37
+ hps = utils.get_hparams_from_file(f"checkpoints/{path}/config.json")
38
+ SVCVITS = SynthesizerTrn(
39
+ hps.data.filter_length // 2 + 1,
40
+ hps.train.segment_size // hps.data.hop_length,
41
+ **hps.model)
42
+ _ = utils.load_checkpoint(f"checkpoints/{path}/model.pth", SVCVITS, None)
43
+ _ = SVCVITS.eval().to(device)
44
+ for i in SVCVITS.parameters():
45
+ i.requires_grad = False
46
+ test_hidden_unit = torch.rand(1, 50, 256)
47
+ test_lengths = torch.LongTensor([50])
48
+ test_pitch = torch.rand(1, 50)
49
+ test_sid = torch.LongTensor([0])
50
+ input_names = ["hidden_unit", "lengths", "pitch", "sid"]
51
+ output_names = ["audio", ]
52
+ SVCVITS.eval()
53
+ torch.onnx.export(SVCVITS,
54
+ (
55
+ test_hidden_unit.to(device),
56
+ test_lengths.to(device),
57
+ test_pitch.to(device),
58
+ test_sid.to(device)
59
+ ),
60
+ f"checkpoints/{path}/model.onnx",
61
+ dynamic_axes={
62
+ "hidden_unit": [0, 1],
63
+ "pitch": [1]
64
+ },
65
+ do_constant_folding=False,
66
+ opset_version=16,
67
+ verbose=False,
68
+ input_names=input_names,
69
+ output_names=output_names)
70
+
71
+
72
+ if __name__ == '__main__':
73
+ main(False,True)
preprocess_flist_config.py ADDED
@@ -0,0 +1,67 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import argparse
3
+ import re
4
+
5
+ from tqdm import tqdm
6
+ from random import shuffle
7
+ import json
8
+
9
+ config_template = json.load(open("configs/config.json"))
10
+
11
+ pattern = re.compile(r'^[\.a-zA-Z0-9_\/]+$')
12
+
13
+ if __name__ == "__main__":
14
+ parser = argparse.ArgumentParser()
15
+ parser.add_argument("--train_list", type=str, default="./filelists/train.txt", help="path to train list")
16
+ parser.add_argument("--val_list", type=str, default="./filelists/val.txt", help="path to val list")
17
+ parser.add_argument("--test_list", type=str, default="./filelists/test.txt", help="path to test list")
18
+ parser.add_argument("--source_dir", type=str, default="./dataset/44k", help="path to source dir")
19
+ args = parser.parse_args()
20
+
21
+ train = []
22
+ val = []
23
+ test = []
24
+ idx = 0
25
+ spk_dict = {}
26
+ spk_id = 0
27
+ for speaker in tqdm(os.listdir(args.source_dir)):
28
+ spk_dict[speaker] = spk_id
29
+ spk_id += 1
30
+ wavs = ["/".join([args.source_dir, speaker, i]) for i in os.listdir(os.path.join(args.source_dir, speaker))]
31
+ for wavpath in wavs:
32
+ if not pattern.match(wavpath):
33
+ print(f"warning:文件名{wavpath}中包含非字母数字下划线,可能会导致错误。(也可能不会)")
34
+ if len(wavs) < 10:
35
+ print(f"warning:{speaker}数据集数量小于10条,请补充数据")
36
+ wavs = [i for i in wavs if i.endswith("wav")]
37
+ shuffle(wavs)
38
+ train += wavs[2:-2]
39
+ val += wavs[:2]
40
+ test += wavs[-2:]
41
+
42
+ shuffle(train)
43
+ shuffle(val)
44
+ shuffle(test)
45
+
46
+ print("Writing", args.train_list)
47
+ with open(args.train_list, "w") as f:
48
+ for fname in tqdm(train):
49
+ wavpath = fname
50
+ f.write(wavpath + "\n")
51
+
52
+ print("Writing", args.val_list)
53
+ with open(args.val_list, "w") as f:
54
+ for fname in tqdm(val):
55
+ wavpath = fname
56
+ f.write(wavpath + "\n")
57
+
58
+ print("Writing", args.test_list)
59
+ with open(args.test_list, "w") as f:
60
+ for fname in tqdm(test):
61
+ wavpath = fname
62
+ f.write(wavpath + "\n")
63
+
64
+ config_template["spk"] = spk_dict
65
+ print("Writing configs/config.json")
66
+ with open("configs/config.json", "w") as f:
67
+ json.dump(config_template, f, indent=2)
preprocess_hubert_f0.py ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import multiprocessing
3
+ import os
4
+ import argparse
5
+ from random import shuffle
6
+
7
+ import torch
8
+ from glob import glob
9
+ from tqdm import tqdm
10
+
11
+ import utils
12
+ import logging
13
+ logging.getLogger('numba').setLevel(logging.WARNING)
14
+ import librosa
15
+ import numpy as np
16
+
17
+ hps = utils.get_hparams_from_file("configs/config.json")
18
+ sampling_rate = hps.data.sampling_rate
19
+ hop_length = hps.data.hop_length
20
+
21
+
22
+ def process_one(filename, hmodel):
23
+ # print(filename)
24
+ wav, sr = librosa.load(filename, sr=sampling_rate)
25
+ soft_path = filename + ".soft.pt"
26
+ if not os.path.exists(soft_path):
27
+ devive = torch.device("cuda" if torch.cuda.is_available() else "cpu")
28
+ wav16k = librosa.resample(wav, orig_sr=sampling_rate, target_sr=16000)
29
+ wav16k = torch.from_numpy(wav16k).to(devive)
30
+ c = utils.get_hubert_content(hmodel, wav_16k_tensor=wav16k)
31
+ torch.save(c.cpu(), soft_path)
32
+ f0_path = filename + ".f0.npy"
33
+ if not os.path.exists(f0_path):
34
+ f0 = utils.compute_f0_dio(wav, sampling_rate=sampling_rate, hop_length=hop_length)
35
+ np.save(f0_path, f0)
36
+
37
+
38
+ def process_batch(filenames):
39
+ print("Loading hubert for content...")
40
+ device = "cuda" if torch.cuda.is_available() else "cpu"
41
+ hmodel = utils.get_hubert_model().to(device)
42
+ print("Loaded hubert.")
43
+ for filename in tqdm(filenames):
44
+ process_one(filename, hmodel)
45
+
46
+
47
+ if __name__ == "__main__":
48
+ parser = argparse.ArgumentParser()
49
+ parser.add_argument("--in_dir", type=str, default="dataset/44k", help="path to input dir")
50
+
51
+ args = parser.parse_args()
52
+ filenames = glob(f'{args.in_dir}/*/*.wav', recursive=True) # [:10]
53
+ shuffle(filenames)
54
+ multiprocessing.set_start_method('spawn')
55
+
56
+ num_processes = 1
57
+ chunk_size = int(math.ceil(len(filenames) / num_processes))
58
+ chunks = [filenames[i:i + chunk_size] for i in range(0, len(filenames), chunk_size)]
59
+ print([len(c) for c in chunks])
60
+ processes = [multiprocessing.Process(target=process_batch, args=(chunk,)) for chunk in chunks]
61
+ for p in processes:
62
+ p.start()
requirements.txt ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Flask
2
+ Flask_Cors
3
+ gradio
4
+ numpy
5
+ playsound
6
+ pydub
7
+ requests
8
+ scipy
9
+ sounddevice
10
+ SoundFile
11
+ starlette
12
+ torch
13
+ torchaudio
14
+ tqdm
15
+ scikit-maad
16
+ praat-parselmouth
17
+ onnx
18
+ onnxsim
19
+ onnxoptimizer
20
+ fairseq
21
+ librosa
resample.py ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import argparse
3
+ import librosa
4
+ import numpy as np
5
+ from multiprocessing import Pool, cpu_count
6
+ from scipy.io import wavfile
7
+ from tqdm import tqdm
8
+
9
+
10
+ def process(item):
11
+ spkdir, wav_name, args = item
12
+ # speaker 's5', 'p280', 'p315' are excluded,
13
+ speaker = spkdir.replace("\\", "/").split("/")[-1]
14
+ wav_path = os.path.join(args.in_dir, speaker, wav_name)
15
+ if os.path.exists(wav_path) and '.wav' in wav_path:
16
+ os.makedirs(os.path.join(args.out_dir2, speaker), exist_ok=True)
17
+ wav, sr = librosa.load(wav_path, None)
18
+ wav, _ = librosa.effects.trim(wav, top_db=20)
19
+ peak = np.abs(wav).max()
20
+ if peak > 1.0:
21
+ wav = 0.98 * wav / peak
22
+ wav2 = librosa.resample(wav, orig_sr=sr, target_sr=args.sr2)
23
+ wav2 /= max(wav2.max(), -wav2.min())
24
+ save_name = wav_name
25
+ save_path2 = os.path.join(args.out_dir2, speaker, save_name)
26
+ wavfile.write(
27
+ save_path2,
28
+ args.sr2,
29
+ (wav2 * np.iinfo(np.int16).max).astype(np.int16)
30
+ )
31
+
32
+
33
+
34
+ if __name__ == "__main__":
35
+ parser = argparse.ArgumentParser()
36
+ parser.add_argument("--sr2", type=int, default=44100, help="sampling rate")
37
+ parser.add_argument("--in_dir", type=str, default="./dataset_raw", help="path to source dir")
38
+ parser.add_argument("--out_dir2", type=str, default="./dataset/44k", help="path to target dir")
39
+ args = parser.parse_args()
40
+ processs = cpu_count()-2 if cpu_count() >4 else 1
41
+ pool = Pool(processes=processs)
42
+
43
+ for speaker in os.listdir(args.in_dir):
44
+ spk_dir = os.path.join(args.in_dir, speaker)
45
+ if os.path.isdir(spk_dir):
46
+ print(spk_dir)
47
+ for _ in tqdm(pool.imap_unordered(process, [(spk_dir, i, args) for i in os.listdir(spk_dir) if i.endswith("wav")])):
48
+ pass