[WIP] add EnCodec model (#23655)

* boilerplate stuff

* messing around with the feature extractor

* fix feature extractor

* unit tests for feature extractor

* rename speech to audio

* quick-and-dirty import of Meta's code

* import weights (sort of)

* cleaning up

* more cleaning up

* move encoder/decoder args into config

* cleanup model

* rename EnCodec -> Encodec

* RVQ parameters in config

* add slow test

* add lstm init and test_init

* Add save & load

* finish EncodecModel

* remove decoder_input_values as they are ont used anywhere (not removed from doc yet)

* fix test feature extraction model name

* Add better slow test

* Fix tests

* some fixup and cleaning

* Improve further

* cleaning up quantizer

* fix up conversion script

* test don't pass, _encode_fram does not work

* update tests with output per encode and decode

* more cleanup

* rename _codebook

* remove old config cruft

* ratios & hop_length

* use ModuleList instead of Sequential

* clean up resnet block

* update types

* update tests

* fixup

* quick cleanup

* fix padding

* more styl,ing

* add patrick feedback

* fix copies

* fixup

* fix lstm

* fix shape issues

* fixup

* rename conv layers

* fixup

* fix decoding

* small conv refactoring

* remove norm_params

* simplify conv layers

* rename conv layers

* stuff

* Clean up

* Add padding logic

use padding mask

small conv refactoring

remove norm_params

simplify conv layers

rename conv layers

stuff

add batched test

update

Clean up

merge and update for padding

fix padding

fixup

* clean up more

* clean up more

* More clean ups

* cleanup convolutions

* typo

* fix typos

* fixup

* build PR doc?

* start refactoring docstring

* fix don't pad when no strid and chunk

* update docstring

* update docstring

* nits

* update going to lunch

* update config and model

* fix broken testse (becaue of the config changes)

* fix scale computation

* fixu[

* only return dict if speciefied or if config returns it

* remove todos

* update defaults in config

* update conversion script

* fix doctest

* more docstring + fixup

* nits on batched_tests

* more nits

* Apply suggestions from code review
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* update basxed on review

* fix update

* updaet tests

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* fixup

* add overlap and chunl_length_s

* cleanup feature extraction

* teste edge cases truncation and padding

* correct processor values

* update config encodec, nits

* fix tests

* fixup

* fix 24Hz test

* elle tests are green

* fix fixup

* Apply suggestions from code review

* revert readme changes

* fixup

* add example

* use facebook checkpoints

* fix typo

* no pipeline tests

* use slef.pad everywhere we can

* Apply suggestions from code review
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* update based on review

* update

* update mdx

* fix bug and tests

* fixup

* fix doctest

* remove comment

* more nits

* add more coverage for `test_truncation_and_padding`

* fixup

* add last test

* fix text

* nits

* Update tests/models/encodec/test_modeling_encodec.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* take care of the last comments

* typo

* fix test

* nits

* fixup

* Update src/transformers/models/encodec/feature_extraction_encodec.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

---------
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: arthur.zucker@gmail.com <arthur.zucker@gmail.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

README.md

@@ -346,6 +346,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (from Meta AI) released with the paper [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu.
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.

README_es.md

@@ -321,6 +321,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (from Meta AI) released with the paper [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu.
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.

README_hd.md

@@ -293,6 +293,7 @@ conda install -c huggingface transformers
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (Google रिसर्च/स्टैनफोर्ड यूनिवर्सिटी से) साथ में दिया गया पेपर [इलेक्ट्रा: जेनरेटर के बजाय भेदभाव करने वाले के रूप में टेक्स्ट एन्कोडर्स का पूर्व-प्रशिक्षण] (https://arxiv.org/abs/2003.10555) केविन क्लार्क, मिन्ह-थांग लुओंग, क्वोक वी. ले, क्रिस्टोफर डी. मैनिंग द्वारा पोस्ट किया गया।
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (Meta AI से) Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi. द्वाराअनुसंधान पत्र [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) के साथ जारी किया गया
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (Google रिसर्च से) साथ में दिया गया पेपर [सीक्वेंस जेनरेशन टास्क के लिए प्री-ट्रेंड चेकपॉइंट का इस्तेमाल करना](https:/ /arxiv.org/abs/1907.12461) साशा रोठे, शशि नारायण, अलियाक्सि सेवेरिन द्वारा।
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)**(Baidu से) साथ देने वाला पेपर [ERNIE: एन्हांस्ड रिप्रेजेंटेशन थ्रू नॉलेज इंटीग्रेशन](https://arxiv.org/abs/1904.09223) यू सन, शुओहुआन वांग, युकुन ली, शिकुन फेंग, ज़ुई चेन, हान झांग, शिन तियान, डैनक्सियांग झू, हाओ तियान, हुआ वू द्वारा पोस्ट किया गया।
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (Baidu से) Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang. द्वाराअनुसंधान पत्र [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) के साथ जारी किया गया

README_ja.md

@@ -355,6 +355,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (Snap Research から) Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren. から公開された研究論文 [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191)
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (Google Research/Stanford University から) Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning から公開された研究論文: [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555)
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (Meta AI から) Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi. から公開された研究論文 [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438)
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (Google Research から) Sascha Rothe, Shashi Narayan, Aliaksei Severyn から公開された研究論文: [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461)
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (Baidu から) Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu から公開された研究論文: [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223)
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (Baidu から) Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang. から公開された研究論文 [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674)

README_ko.md

@@ -270,6 +270,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (Google Research/Stanford University 에서) Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning 의 [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) 논문과 함께 발표했습니다.
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (Meta AI 에서 제공)은 Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.의 [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438)논문과 함께 발표했습니다.
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (Google Research 에서) Sascha Rothe, Shashi Narayan, Aliaksei Severyn 의 [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) 논문과 함께 발표했습니다.
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (Baidu 에서) Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu 의 [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) 논문과 함께 발표했습니다.
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (Baidu 에서 제공)은 Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.의 [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674)논문과 함께 발표했습니다.

README_zh-hans.md

@@ -294,6 +294,7 @@ conda install -c huggingface transformers
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (来自 Snap Research) 伴随论文 [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) 由 Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren 发布。
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (来自 Google Research/Stanford University) 伴随论文 [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) 由 Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning 发布。
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (来自 Meta AI) 伴随论文 [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) 由 Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi 发布。
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (来自 Google Research) 伴随论文 [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) 由 Sascha Rothe, Shashi Narayan, Aliaksei Severyn 发布。
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (来自 Baidu) 伴随论文 [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu 发布。
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (来自 Baidu) 伴随论文 [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) 由 Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang 发布。

README_zh-hant.md

@@ -306,6 +306,7 @@ conda install -c huggingface transformers
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (from Meta AI) released with the paper [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu.
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.

docs/source/en/_toctree.yml

@@ -541,6 +541,8 @@
         title: Audio Spectrogram Transformer
       - local: model_doc/clap
         title: CLAP
+      - local: model_doc/encodec
+        title: EnCodec
       - local: model_doc/hubert
         title: Hubert
       - local: model_doc/mctct

docs/source/en/index.mdx

@@ -107,6 +107,7 @@ The documentation is organized into five sections:
 1. **[EfficientFormer](model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
 1. **[EfficientNet](model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
+1. **[EnCodec](model_doc/encodec)** (from Meta AI) released with the paper [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
 1. **[EncoderDecoder](model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
 1. **[ERNIE](model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu.
 1. **[ErnieM](model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.
@@ -318,6 +319,7 @@ Flax), PyTorch, and/or TensorFlow.
 |        EfficientFormer        |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |         EfficientNet          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            ELECTRA            |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|            EnCodec            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |        Encoder decoder        |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
 |             ERNIE             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            ErnieM             |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/encodec.mdx

+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# EnCodec
+
+## Overview
+
+The EnCodec neural codec model was proposed in [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
+
+The abstract from the paper is the following:
+
+*We introduce a state-of-the-art real-time, high-fidelity, audio codec leveraging neural networks. It consists in a streaming encoder-decoder architecture with quantized latent space trained in an end-to-end fashion. We simplify and speed-up the training by using a single multiscale spectrogram adversary that efficiently reduces artifacts and produce high-quality samples. We introduce a novel loss balancer mechanism to stabilize training: the weight of a loss now defines the fraction of the overall gradient it should represent, thus decoupling the choice of this hyper-parameter from the typical scale of the loss. Finally, we study how lightweight Transformer models can be used to further compress the obtained representation by up to 40%, while staying faster than real time. We provide a detailed description of the key design choices of the proposed model including: training objective, architectural changes and a study of various perceptual loss functions. We present an extensive subjective evaluation (MUSHRA tests) together with an ablation study for a range of bandwidths and audio domains, including speech, noisy-reverberant speech, and music. Our approach is superior to the baselines methods across all evaluated settings, considering both 24 kHz monophonic and 48 kHz stereophonic audio.*
+
+This model was contributed by [Matthijs](https://huggingface.co/Matthijs), [Patrick Von Platen](https://huggingface.co/patrickvonplaten) and [Arthur Zucker](https://huggingface.co/ArthurZ). 
+The original code can be found [here](https://github.com/facebookresearch/encodec).
+Here is a quick example of how to encode and decode an audio using this model:
+
+```python 
+>>> from datasets import load_dataset, Audio
+>>> from transformers import EncodecModel, AutoProcessor
+>>> librispeech_dummy = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+>>> model = EncodecModel.from_pretrained("facebook/encodec_24khz")
+>>> processor = AutoProcessor.from_pretrained("facebook/encodec_24khz")
+>>> librispeech_dummy = librispeech_dummy.cast_column("audio", Audio(sampling_rate=processor.sampling_rate))
+>>> audio_sample = librispeech_dummy[-1]["audio"]["array"]
+>>> inputs = processor(raw_audio=audio_sample, sampling_rate=processor.sampling_rate, return_tensors="pt")
+
+>>> encoder_outputs = model.encode(inputs["input_values"], inputs["padding_mask"])
+>>> audio_values = model.decode(encoder_outputs.audio_codes, encoder_outputs.audio_scales, inputs["padding_mask"])[0]
+>>> # or the equivalent with a forward pass
+>>> audio_values = model(inputs["input_values"], inputs["padding_mask"]).audio_values
+```
+
+
+## EncodecConfig
+
+[[autodoc]] EncodecConfig
+
+## EncodecFeatureExtractor
+
+[[autodoc]] EncodecFeatureExtractor
+    - __call__
+
+## EncodecModel
+
+[[autodoc]] EncodecModel
+    - decode
+    - encode
+    - forward

src/transformers/__init__.py

@@ -281,6 +281,10 @@ _import_structure = {
     "models.efficientformer": ["EFFICIENTFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "EfficientFormerConfig"],
     "models.efficientnet": ["EFFICIENTNET_PRETRAINED_CONFIG_ARCHIVE_MAP", "EfficientNetConfig"],
     "models.electra": ["ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP", "ElectraConfig", "ElectraTokenizer"],
+    "models.encodec": [
+        "ENCODEC_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "EncodecConfig",
+    ],
     "models.encoder_decoder": ["EncoderDecoderConfig"],
     "models.ernie": [
         "ERNIE_PRETRAINED_CONFIG_ARCHIVE_MAP",
@@ -825,6 +829,7 @@ except OptionalDependencyNotAvailable:
     ]
 else:
     _import_structure["models.audio_spectrogram_transformer"].append("ASTFeatureExtractor")
+    _import_structure["models.encodec"].append("EncodecFeatureExtractor")
     _import_structure["models.mctct"].append("MCTCTFeatureExtractor")
     _import_structure["models.speech_to_text"].append("Speech2TextFeatureExtractor")
     _import_structure["models.speecht5"].append("SpeechT5FeatureExtractor")
@@ -1568,6 +1573,13 @@ else:
             "load_tf_weights_in_electra",
         ]
     )
+    _import_structure["models.encodec"].extend(
+        [
+            "ENCODEC_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "EncodecModel",
+            "EncodecPreTrainedModel",
+        ]
+    )
     _import_structure["models.encoder_decoder"].append("EncoderDecoderModel")
     _import_structure["models.ernie"].extend(
         [
@@ -4100,6 +4112,10 @@ if TYPE_CHECKING:
     from .models.efficientformer import EFFICIENTFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, EfficientFormerConfig
     from .models.efficientnet import EFFICIENTNET_PRETRAINED_CONFIG_ARCHIVE_MAP, EfficientNetConfig
     from .models.electra import ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP, ElectraConfig, ElectraTokenizer
+    from .models.encodec import (
+        ENCODEC_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        EncodecConfig,
+    )
     from .models.encoder_decoder import EncoderDecoderConfig
     from .models.ernie import ERNIE_PRETRAINED_CONFIG_ARCHIVE_MAP, ErnieConfig
     from .models.ernie_m import ERNIE_M_PRETRAINED_CONFIG_ARCHIVE_MAP, ErnieMConfig
@@ -4598,6 +4614,7 @@ if TYPE_CHECKING:
         from .utils.dummy_speech_objects import *
     else:
         from .models.audio_spectrogram_transformer import ASTFeatureExtractor
+        from .models.encodec import EncodecFeatureExtractor
         from .models.mctct import MCTCTFeatureExtractor
         from .models.speech_to_text import Speech2TextFeatureExtractor
         from .models.speecht5 import SpeechT5FeatureExtractor
@@ -5210,6 +5227,11 @@ if TYPE_CHECKING:
             ElectraPreTrainedModel,
             load_tf_weights_in_electra,
         )
+        from .models.encodec import (
+            ENCODEC_PRETRAINED_MODEL_ARCHIVE_LIST,
+            EncodecModel,
+            EncodecPreTrainedModel,
+        )
         from .models.encoder_decoder import EncoderDecoderModel
         from .models.ernie import (
             ERNIE_PRETRAINED_MODEL_ARCHIVE_LIST,

src/transformers/models/__init__.py

@@ -72,6 +72,7 @@ from . import (
     efficientformer,
     efficientnet,
     electra,
+    encodec,
     encoder_decoder,
     ernie,
     ernie_m,

src/transformers/models/auto/configuration_auto.py

@@ -80,6 +80,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("efficientformer", "EfficientFormerConfig"),
         ("efficientnet", "EfficientNetConfig"),
         ("electra", "ElectraConfig"),
+        ("encodec", "EncodecConfig"),
         ("encoder-decoder", "EncoderDecoderConfig"),
         ("ernie", "ErnieConfig"),
         ("ernie_m", "ErnieMConfig"),
@@ -273,6 +274,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("efficientformer", "EFFICIENTFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("efficientnet", "EFFICIENTNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("electra", "ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("encodec", "ENCODEC_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("ernie", "ERNIE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("ernie_m", "ERNIE_M_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("esm", "ESM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -461,6 +463,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("efficientformer", "EfficientFormer"),
         ("efficientnet", "EfficientNet"),
         ("electra", "ELECTRA"),
+        ("encodec", "EnCodec"),
         ("encoder-decoder", "Encoder decoder"),
         ("ernie", "ERNIE"),
         ("ernie_m", "ErnieM"),

src/transformers/models/auto/feature_extraction_auto.py

@@ -54,6 +54,7 @@ FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict(
         ("dinat", "ViTFeatureExtractor"),
         ("donut-swin", "DonutFeatureExtractor"),
         ("dpt", "DPTFeatureExtractor"),
+        ("encodec", "EncodecFeatureExtractor"),
         ("flava", "FlavaFeatureExtractor"),
         ("glpn", "GLPNFeatureExtractor"),
         ("groupvit", "CLIPFeatureExtractor"),

src/transformers/models/auto/modeling_auto.py

@@ -79,6 +79,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("efficientformer", "EfficientFormerModel"),
         ("efficientnet", "EfficientNetModel"),
         ("electra", "ElectraModel"),
+        ("encodec", "EncodecModel"),
         ("ernie", "ErnieModel"),
         ("ernie_m", "ErnieMModel"),
         ("esm", "EsmModel"),

src/transformers/models/encodec/__init__.py

+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_encodec": [
+        "ENCODEC_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "EncodecConfig",
+    ],
+    "feature_extraction_encodec": ["EncodecFeatureExtractor"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_encodec"] = [
+        "ENCODEC_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "EncodecModel",
+        "EncodecPreTrainedModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_encodec import (
+        ENCODEC_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        EncodecConfig,
+    )
+    from .feature_extraction_encodec import EncodecFeatureExtractor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_encodec import (
+            ENCODEC_PRETRAINED_MODEL_ARCHIVE_LIST,
+            EncodecModel,
+            EncodecPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/encodec/configuration_encodec.py

+# coding=utf-8
+# Copyright 2023 Meta Platforms, Inc. and affiliates, and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" EnCodec model configuration"""
+
+
+import math
+from typing import Optional
+
+import numpy as np
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+ENCODEC_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/encodec_24khz": "https://huggingface.co/facebook/encodec_24khz/resolve/main/config.json",
+    "facebook/encodec_48khz": "https://huggingface.co/facebook/encodec_48khz/resolve/main/config.json",
+}
+
+
+class EncodecConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of an [`EncodecModel`]. It is used to instantiate a
+    Encodec model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the
+    [facebook/encodec_24khz](https://huggingface.co/facebook/encodec_24khz) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        target_bandwidths (`List[float]`, *optional*, defaults to `[1.5, 3.0, 6.0, 12.0, 24.0]`):
+            The range of diffent bandwiths the model can encode audio with.
+        sampling_rate (`int`, *optional*, defaults to 24000):
+            The sampling rate at which the audio waveform should be digitalized expressed in hertz (Hz).
+        audio_channels (`int`, *optional*, defaults to 1):
+            Number of channels in the audio data. Either 1 for mono or 2 for stereo.
+        normalize (`bool`, *optional*, defaults to `False`):
+            Whether the audio shall be normalized when passed.
+        chunk_length_s (`float`, *optional*):
+            If defined the audio is pre-processed into chunks of lengths `chunk_length_s` and then encoded.
+        overlap (`float`, *optional*):
+            Defines the overlap between each chunk. It is used to compute the `chunk_stride` using the following
+            formulae : `int((1.0 - self.overlap) * self.chunk_length)`.
+        hidden_size (`int`, *optional*, defaults to 128):
+            Intermediate representation dimension.
+        num_filters (`int`, *optional*, defaults to 32):
+            Number of convolution kernels of first `EncodecConv1d` down sampling layer.
+        num_residual_layers (`int`,  *optional*, defaults to 1):
+            Number of residual layers.
+        upsampling_ratios (`Sequence[int]` , *optional*, defaults to `[8, 5, 4, 2]`):
+            Kernel size and stride ratios. The encoder uses downsampling ratios instead of upsampling ratios, hence it
+            will use the ratios in the reverse order to the ones specified here that must match the decoder order.
+        norm_type (`str`, *optional*, defaults to `"weight_norm"`):
+            Normalization method. Should be in `["weight_norm", "time_group_norm"]`
+        kernel_size (`int`, *optional*, defaults to 7):
+            Kernel size for the initial convolution.
+        last_kernel_size (`int`, *optional*, defaults to 7):
+            Kernel size for the last convolution layer.
+        residual_kernel_size (`int`, *optional*, defaults to 3):
+            Kernel size for the residual layers.
+        dilation_growth_rate (`int`, *optional*, defaults to 2):
+            How much to increase the dilation with each layer.
+        use_causal_conv (`bool`, *optional*, defaults to `True`):
+            Whether to use fully causal convolution.
+        pad_mode (`str`, *optional*, defaults to `"reflect"`):
+            Padding mode for the convolutions.
+        compress (`int`, *optional*, defaults to 2):
+            Reduced dimensionality in residual branches (from Demucs v3).
+        num_lstm_layers (`int`, *optional*, defaults to 2):
+            Number of LSTM layers at the end of the encoder.
+        trim_right_ratio (`float`, *optional*, defaults to 1.0):
+            Ratio for trimming at the right of the transposed convolution under the `use_causal_conv = True` setup. If
+            equal to 1.0, it means that all the trimming is done at the right.
+        codebook_size (`int`, *optional*, defaults to 1024):
+            Number of discret codes that make up VQVAE.
+        codebook_dim (`int`, *optional*):
+            Dimension of the codebook vectors. If not defined, uses `hidden_size`.
+
+    Example:
+
+    ```python
+    >>> from transformers import EncodecModel, EncodecConfig
+
+    >>> # Initializing a "facebook/encodec_24khz" style configuration
+    >>> configuration = EncodecConfig()
+
+    >>> # Initializing a model (with random weights) from the "facebook/encodec_24khz" style configuration
+    >>> model = EncodecModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "encodec"
+
+    def __init__(
+        self,
+        target_bandwidths=[1.5, 3.0, 6.0, 12.0, 24.0],
+        sampling_rate=24_000,
+        audio_channels=1,
+        normalize=False,
+        chunk_length_s=None,
+        overlap=None,
+        hidden_size=128,
+        num_filters=32,
+        num_residual_layers=1,
+        upsampling_ratios=[8, 5, 4, 2],
+        norm_type="weight_norm",
+        kernel_size=7,
+        last_kernel_size=7,
+        residual_kernel_size=3,
+        dilation_growth_rate=2,
+        use_causal_conv=True,
+        pad_mode="reflect",
+        compress=2,
+        num_lstm_layers=2,
+        trim_right_ratio=1.0,
+        codebook_size=1024,
+        codebook_dim=None,
+        **kwargs,
+    ):
+        self.target_bandwidths = target_bandwidths
+        self.sampling_rate = sampling_rate
+        self.audio_channels = audio_channels
+        self.normalize = normalize
+        self.chunk_length_s = chunk_length_s
+        self.overlap = overlap
+        self.hidden_size = hidden_size
+        self.num_filters = num_filters
+        self.num_residual_layers = num_residual_layers
+        self.upsampling_ratios = upsampling_ratios
+        self.norm_type = norm_type
+        self.kernel_size = kernel_size
+        self.last_kernel_size = last_kernel_size
+        self.residual_kernel_size = residual_kernel_size
+        self.dilation_growth_rate = dilation_growth_rate
+        self.use_causal_conv = use_causal_conv
+        self.pad_mode = pad_mode
+        self.compress = compress
+        self.num_lstm_layers = num_lstm_layers
+        self.trim_right_ratio = trim_right_ratio
+        self.codebook_size = codebook_size
+        self.codebook_dim = codebook_dim if codebook_dim is not None else hidden_size
+
+        if self.norm_type not in ["weight_norm", "time_group_norm"]:
+            raise ValueError(
+                f'self.norm_type must be one of `"weight_norm"`, `"time_group_norm"`), got {self.norm_type}'
+            )
+
+        super().__init__(**kwargs)
+
+    # This is a property because you might want to change the chunk_length_s on the fly
+    @property
+    def chunk_length(self) -> Optional[int]:
+        if self.chunk_length_s is None:
+            return None
+        else:
+            return int(self.chunk_length_s * self.sampling_rate)
+
+    # This is a property because you might want to change the chunk_length_s on the fly
+    @property
+    def chunk_stride(self) -> Optional[int]:
+        if self.chunk_length_s is None or self.overlap is None:
+            return None
+        else:
+            return max(1, int((1.0 - self.overlap) * self.chunk_length))
+
+    @property
+    def frame_rate(self) -> int:
+        hop_length = np.prod(self.upsampling_ratios)
+        return math.ceil(self.sampling_rate / hop_length)
+
+    @property
+    def num_quantizers(self) -> int:
+        return int(1000 * self.target_bandwidths[-1] // (self.frame_rate * 10))

src/transformers/models/encodec/convert_encodec_checkpoint_to_pytorch.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert EnCodec checkpoints."""
+
+import argparse
+
+import torch
+
+from transformers import (
+    EncodecConfig,
+    EncodecFeatureExtractor,
+    EncodecModel,
+    logging,
+)
+
+
+# checkpoints downloaded from:
+# https://dl.fbaipublicfiles.com/encodec/v0/encodec_24khz-d7cc33bc.th
+# https://dl.fbaipublicfiles.com/encodec/v0/encodec_48khz-7e698e3e.th
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger("transformers.models.encodec")
+
+MAPPING_QUANTIZER = {
+    "quantizer.vq.layers.*._codebook.inited": "quantizer.layers.*.codebook.inited",
+    "quantizer.vq.layers.*._codebook.cluster_size": "quantizer.layers.*.codebook.cluster_size",
+    "quantizer.vq.layers.*._codebook.embed": "quantizer.layers.*.codebook.embed",
+    "quantizer.vq.layers.*._codebook.embed_avg": "quantizer.layers.*.codebook.embed_avg",
+}
+MAPPING_ENCODER = {
+    "encoder.model.0.conv.conv": "encoder.layers.0.conv",
+    "encoder.model.1.block.1.conv.conv": "encoder.layers.1.block.1.conv",
+    "encoder.model.1.block.3.conv.conv": "encoder.layers.1.block.3.conv",
+    "encoder.model.1.shortcut.conv.conv": "encoder.layers.1.shortcut.conv",
+    "encoder.model.3.conv.conv": "encoder.layers.3.conv",
+    "encoder.model.4.block.1.conv.conv": "encoder.layers.4.block.1.conv",
+    "encoder.model.4.block.3.conv.conv": "encoder.layers.4.block.3.conv",
+    "encoder.model.4.shortcut.conv.conv": "encoder.layers.4.shortcut.conv",
+    "encoder.model.6.conv.conv": "encoder.layers.6.conv",
+    "encoder.model.7.block.1.conv.conv": "encoder.layers.7.block.1.conv",
+    "encoder.model.7.block.3.conv.conv": "encoder.layers.7.block.3.conv",
+    "encoder.model.7.shortcut.conv.conv": "encoder.layers.7.shortcut.conv",
+    "encoder.model.9.conv.conv": "encoder.layers.9.conv",
+    "encoder.model.10.block.1.conv.conv": "encoder.layers.10.block.1.conv",
+    "encoder.model.10.block.3.conv.conv": "encoder.layers.10.block.3.conv",
+    "encoder.model.10.shortcut.conv.conv": "encoder.layers.10.shortcut.conv",
+    "encoder.model.12.conv.conv": "encoder.layers.12.conv",
+    "encoder.model.13.lstm": "encoder.layers.13.lstm",
+    "encoder.model.15.conv.conv": "encoder.layers.15.conv",
+}
+MAPPING_ENCODER_48K = {
+    "encoder.model.0.conv.norm": "encoder.layers.0.norm",
+    "encoder.model.1.block.1.conv.norm": "encoder.layers.1.block.1.norm",
+    "encoder.model.1.block.3.conv.norm": "encoder.layers.1.block.3.norm",
+    "encoder.model.1.shortcut.conv.norm": "encoder.layers.1.shortcut.norm",
+    "encoder.model.3.conv.norm": "encoder.layers.3.norm",
+    "encoder.model.4.block.1.conv.norm": "encoder.layers.4.block.1.norm",
+    "encoder.model.4.block.3.conv.norm": "encoder.layers.4.block.3.norm",
+    "encoder.model.4.shortcut.conv.norm": "encoder.layers.4.shortcut.norm",
+    "encoder.model.6.conv.norm": "encoder.layers.6.norm",
+    "encoder.model.7.block.1.conv.norm": "encoder.layers.7.block.1.norm",
+    "encoder.model.7.block.3.conv.norm": "encoder.layers.7.block.3.norm",
+    "encoder.model.7.shortcut.conv.norm": "encoder.layers.7.shortcut.norm",
+    "encoder.model.9.conv.norm": "encoder.layers.9.norm",
+    "encoder.model.10.block.1.conv.norm": "encoder.layers.10.block.1.norm",
+    "encoder.model.10.block.3.conv.norm": "encoder.layers.10.block.3.norm",
+    "encoder.model.10.shortcut.conv.norm": "encoder.layers.10.shortcut.norm",
+    "encoder.model.12.conv.norm": "encoder.layers.12.norm",
+    "encoder.model.15.conv.norm": "encoder.layers.15.norm",
+}
+MAPPING_DECODER = {
+    "decoder.model.0.conv.conv": "decoder.layers.0.conv",
+    "decoder.model.1.lstm": "decoder.layers.1.lstm",
+    "decoder.model.3.convtr.convtr": "decoder.layers.3.conv",
+    "decoder.model.4.block.1.conv.conv": "decoder.layers.4.block.1.conv",
+    "decoder.model.4.block.3.conv.conv": "decoder.layers.4.block.3.conv",
+    "decoder.model.4.shortcut.conv.conv": "decoder.layers.4.shortcut.conv",
+    "decoder.model.6.convtr.convtr": "decoder.layers.6.conv",
+    "decoder.model.7.block.1.conv.conv": "decoder.layers.7.block.1.conv",
+    "decoder.model.7.block.3.conv.conv": "decoder.layers.7.block.3.conv",
+    "decoder.model.7.shortcut.conv.conv": "decoder.layers.7.shortcut.conv",
+    "decoder.model.9.convtr.convtr": "decoder.layers.9.conv",
+    "decoder.model.10.block.1.conv.conv": "decoder.layers.10.block.1.conv",
+    "decoder.model.10.block.3.conv.conv": "decoder.layers.10.block.3.conv",
+    "decoder.model.10.shortcut.conv.conv": "decoder.layers.10.shortcut.conv",
+    "decoder.model.12.convtr.convtr": "decoder.layers.12.conv",
+    "decoder.model.13.block.1.conv.conv": "decoder.layers.13.block.1.conv",
+    "decoder.model.13.block.3.conv.conv": "decoder.layers.13.block.3.conv",
+    "decoder.model.13.shortcut.conv.conv": "decoder.layers.13.shortcut.conv",
+    "decoder.model.15.conv.conv": "decoder.layers.15.conv",
+}
+MAPPING_DECODER_48K = {
+    "decoder.model.0.conv.norm": "decoder.layers.0.norm",
+    "decoder.model.3.convtr.norm": "decoder.layers.3.norm",
+    "decoder.model.4.block.1.conv.norm": "decoder.layers.4.block.1.norm",
+    "decoder.model.4.block.3.conv.norm": "decoder.layers.4.block.3.norm",
+    "decoder.model.4.shortcut.conv.norm": "decoder.layers.4.shortcut.norm",
+    "decoder.model.6.convtr.norm": "decoder.layers.6.norm",
+    "decoder.model.7.block.1.conv.norm": "decoder.layers.7.block.1.norm",
+    "decoder.model.7.block.3.conv.norm": "decoder.layers.7.block.3.norm",
+    "decoder.model.7.shortcut.conv.norm": "decoder.layers.7.shortcut.norm",
+    "decoder.model.9.convtr.norm": "decoder.layers.9.norm",
+    "decoder.model.10.block.1.conv.norm": "decoder.layers.10.block.1.norm",
+    "decoder.model.10.block.3.conv.norm": "decoder.layers.10.block.3.norm",
+    "decoder.model.10.shortcut.conv.norm": "decoder.layers.10.shortcut.norm",
+    "decoder.model.12.convtr.norm": "decoder.layers.12.norm",
+    "decoder.model.13.block.1.conv.norm": "decoder.layers.13.block.1.norm",
+    "decoder.model.13.block.3.conv.norm": "decoder.layers.13.block.3.norm",
+    "decoder.model.13.shortcut.conv.norm": "decoder.layers.13.shortcut.norm",
+    "decoder.model.15.conv.norm": "decoder.layers.15.norm",
+}
+MAPPING_24K = {
+    **MAPPING_QUANTIZER,
+    **MAPPING_ENCODER,
+    **MAPPING_DECODER,
+}
+MAPPING_48K = {
+    **MAPPING_QUANTIZER,
+    **MAPPING_ENCODER,
+    **MAPPING_ENCODER_48K,
+    **MAPPING_DECODER,
+    **MAPPING_DECODER_48K,
+}
+TOP_LEVEL_KEYS = []
+IGNORE_KEYS = []
+
+
+def set_recursively(hf_pointer, key, value, full_name, weight_type):
+    for attribute in key.split("."):
+        hf_pointer = getattr(hf_pointer, attribute)
+
+    if weight_type is not None:
+        hf_shape = getattr(hf_pointer, weight_type).shape
+    else:
+        hf_shape = hf_pointer.shape
+
+    if hf_shape != value.shape:
+        raise ValueError(
+            f"Shape of hf {key + '.' + weight_type if weight_type is not None else ''} is {hf_shape}, but should be"
+            f" {value.shape} for {full_name}"
+        )
+
+    if weight_type == "weight":
+        hf_pointer.weight.data = value
+    elif weight_type == "weight_g":
+        hf_pointer.weight_g.data = value
+    elif weight_type == "weight_v":
+        hf_pointer.weight_v.data = value
+    elif weight_type == "bias":
+        hf_pointer.bias.data = value
+    elif weight_type == "running_mean":
+        hf_pointer.running_mean.data = value
+    elif weight_type == "running_var":
+        hf_pointer.running_var.data = value
+    elif weight_type == "num_batches_tracked":
+        hf_pointer.num_batches_tracked.data = value
+    elif weight_type == "weight_ih_l0":
+        hf_pointer.weight_ih_l0.data = value
+    elif weight_type == "weight_hh_l0":
+        hf_pointer.weight_hh_l0.data = value
+    elif weight_type == "bias_ih_l0":
+        hf_pointer.bias_ih_l0.data = value
+    elif weight_type == "bias_hh_l0":
+        hf_pointer.bias_hh_l0.data = value
+    elif weight_type == "weight_ih_l1":
+        hf_pointer.weight_ih_l1.data = value
+    elif weight_type == "weight_hh_l1":
+        hf_pointer.weight_hh_l1.data = value
+    elif weight_type == "bias_ih_l1":
+        hf_pointer.bias_ih_l1.data = value
+    elif weight_type == "bias_hh_l1":
+        hf_pointer.bias_hh_l1.data = value
+    else:
+        hf_pointer.data = value
+
+    logger.info(f"{key + ('.' + weight_type if weight_type is not None else '')} was initialized from {full_name}.")
+
+
+def should_ignore(name, ignore_keys):
+    for key in ignore_keys:
+        if key.endswith(".*"):
+            if name.startswith(key[:-1]):
+                return True
+        elif ".*." in key:
+            prefix, suffix = key.split(".*.")
+            if prefix in name and suffix in name:
+                return True
+        elif key in name:
+            return True
+    return False
+
+
+def recursively_load_weights(orig_dict, hf_model, model_name):
+    unused_weights = []
+
+    if model_name == "encodec_24khz":
+        MAPPING = MAPPING_24K
+    elif model_name == "encodec_48khz":
+        MAPPING = MAPPING_48K
+    else:
+        raise ValueError(f"Unsupported model: {model_name}")
+
+    for name, value in orig_dict.items():
+        if should_ignore(name, IGNORE_KEYS):
+            logger.info(f"{name} was ignored")
+            continue
+
+        is_used = False
+        for key, mapped_key in MAPPING.items():
+            if "*" in key:
+                prefix, suffix = key.split(".*.")
+                if prefix in name and suffix in name:
+                    key = suffix
+
+            if key in name:
+                # HACK otherwise .embed gets initialized with .embed_avg too
+                if key.endswith("embed") and name.endswith("embed_avg"):
+                    continue
+
+                is_used = True
+                if "*" in mapped_key:
+                    layer_index = name.split(key)[0].split(".")[-2]
+                    mapped_key = mapped_key.replace("*", layer_index)
+                if "weight_g" in name:
+                    weight_type = "weight_g"
+                elif "weight_v" in name:
+                    weight_type = "weight_v"
+                elif "weight_ih_l0" in name:
+                    weight_type = "weight_ih_l0"
+                elif "weight_hh_l0" in name:
+                    weight_type = "weight_hh_l0"
+                elif "bias_ih_l0" in name:
+                    weight_type = "bias_ih_l0"
+                elif "bias_hh_l0" in name:
+                    weight_type = "bias_hh_l0"
+                elif "weight_ih_l1" in name:
+                    weight_type = "weight_ih_l1"
+                elif "weight_hh_l1" in name:
+                    weight_type = "weight_hh_l1"
+                elif "bias_ih_l1" in name:
+                    weight_type = "bias_ih_l1"
+                elif "bias_hh_l1" in name:
+                    weight_type = "bias_hh_l1"
+                elif "bias" in name:
+                    weight_type = "bias"
+                elif "weight" in name:
+                    weight_type = "weight"
+                elif "running_mean" in name:
+                    weight_type = "running_mean"
+                elif "running_var" in name:
+                    weight_type = "running_var"
+                elif "num_batches_tracked" in name:
+                    weight_type = "num_batches_tracked"
+                else:
+                    weight_type = None
+                set_recursively(hf_model, mapped_key, value, name, weight_type)
+            continue
+        if not is_used:
+            unused_weights.append(name)
+
+    logger.warning(f"Unused weights: {unused_weights}")
+
+
+@torch.no_grad()
+def convert_checkpoint(
+    model_name,
+    checkpoint_path,
+    pytorch_dump_folder_path,
+    config_path=None,
+    repo_id=None,
+):
+    """
+    Copy/paste/tweak model's weights to transformers design.
+    """
+    if config_path is not None:
+        config = EncodecConfig.from_pretrained(config_path)
+    else:
+        config = EncodecConfig()
+
+    if model_name == "encodec_24khz":
+        pass  # config is already correct
+    elif model_name == "encodec_48khz":
+        config.upsampling_ratios = [8, 5, 4, 2]
+        config.target_bandwidths = [3.0, 6.0, 12.0, 24.0]
+        config.sampling_rate = 48_000
+        config.audio_channels = 2
+        config.use_causal_conv = False
+        config.norm_type = "time_group_norm"
+        config.normalize = True
+        config.chunk_length_s = 1.0
+        config.overlap = 0.01
+    else:
+        raise ValueError(f"Unknown model name: {model_name}")
+
+    model = EncodecModel(config)
+
+    feature_extractor = EncodecFeatureExtractor(
+        feature_size=config.audio_channels,
+        sampling_rate=config.sampling_rate,
+        chunk_length_s=config.chunk_length_s,
+        overlap=config.overlap,
+    )
+    feature_extractor.save_pretrained(pytorch_dump_folder_path)
+
+    original_checkpoint = torch.load(checkpoint_path)
+    recursively_load_weights(original_checkpoint, model, model_name)
+    model.save_pretrained(pytorch_dump_folder_path)
+
+    if repo_id:
+        print("Pushing to the hub...")
+        feature_extractor.push_to_hub(repo_id)
+        model.push_to_hub(repo_id)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model",
+        default="encodec_24khz",
+        type=str,
+        help="The model to convert. Should be one of 'encodec_24khz', 'encodec_48khz'.",
+    )
+    parser.add_argument("--checkpoint_path", required=True, default=None, type=str, help="Path to original checkpoint")
+    parser.add_argument("--config_path", default=None, type=str, help="Path to hf config.json of model to convert")
+    parser.add_argument(
+        "--pytorch_dump_folder_path", required=True, default=None, type=str, help="Path to the output PyTorch model."
+    )
+    parser.add_argument(
+        "--push_to_hub", default=None, type=str, help="Where to upload the converted model on the 🤗 hub."
+    )
+
+    args = parser.parse_args()
+    convert_checkpoint(
+        args.model,
+        args.checkpoint_path,
+        args.pytorch_dump_folder_path,
+        args.config_path,
+        args.push_to_hub,
+    )

src/transformers/models/encodec/feature_extraction_encodec.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Feature extractor class for EnCodec."""
+
+from typing import List, Optional, Union
+
+import numpy as np
+
+from ...feature_extraction_sequence_utils import SequenceFeatureExtractor
+from ...feature_extraction_utils import BatchFeature
+from ...utils import PaddingStrategy, TensorType, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class EncodecFeatureExtractor(SequenceFeatureExtractor):
+    r"""
+    Constructs an EnCodec feature extractor.
+
+    This feature extractor inherits from [`~feature_extraction_sequence_utils.SequenceFeatureExtractor`] which contains
+    most of the main methods. Users should refer to this superclass for more information regarding those methods.
+
+    Instantiating a feature extractor with the defaults will yield a similar configuration to that of the
+    [facebook/encodec_24khz](https://huggingface.co/facebook/encodec_24khz) architecture.
+
+    Args:
+        feature_size (`int`, *optional*, defaults to 1):
+            The feature dimension of the extracted features. Use 1 for mono, 2 for stereo.
+        sampling_rate (`int`, *optional*, defaults to 24000):
+            The sampling rate at which the audio waveform should be digitalized expressed in hertz (Hz).
+        padding_value (`float`, *optional*, defaults to 0.0):
+            The value that is used to fill the padding values.
+        chunk_length_s (`float`, *optional*):
+            If defined the audio is pre-processed into chunks of lengths `chunk_length_s` and then encoded.
+        overlap (`float`, *optional*):
+            Defines the overlap between each chunk. It is used to compute the `chunk_stride` using the following
+            formulae : `int((1.0 - self.overlap) * self.chunk_length)`.
+    """
+
+    model_input_names = ["input_values", "padding_mask"]
+
+    def __init__(
+        self,
+        feature_size: int = 1,
+        sampling_rate: int = 24000,
+        padding_value: float = 0.0,
+        chunk_length_s: float = None,
+        overlap: float = None,
+        **kwargs,
+    ):
+        super().__init__(feature_size=feature_size, sampling_rate=sampling_rate, padding_value=padding_value, **kwargs)
+        self.chunk_length_s = chunk_length_s
+        self.overlap = overlap
+
+    # This is a property because you might want to change the chunk_length_s on the fly
+    @property
+    def chunk_length(self) -> Optional[int]:
+        if self.chunk_length_s is None:
+            return None
+        else:
+            return int(self.chunk_length_s * self.sampling_rate)
+
+    # This is a property because you might want to change the chunk_length_s on the fly
+    @property
+    def chunk_stride(self) -> Optional[int]:
+        if self.chunk_length_s is None or self.overlap is None:
+            return None
+        else:
+            return max(1, int((1.0 - self.overlap) * self.chunk_length))
+
+    def __call__(
+        self,
+        raw_audio: Union[np.ndarray, List[float], List[np.ndarray], List[List[float]]],
+        padding: Optional[Union[bool, str, PaddingStrategy]] = None,
+        truncation: Optional[bool] = False,
+        max_length: Optional[int] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        sampling_rate: Optional[int] = None,
+    ) -> BatchFeature:
+        """
+        Main method to featurize and prepare for the model one or several sequence(s).
+
+        Args:
+            raw_audio (`np.ndarray`, `List[float]`, `List[np.ndarray]`, `List[List[float]]`):
+                The sequence or batch of sequences to be processed. Each sequence can be a numpy array, a list of float
+                values, a list of numpy arrays or a list of list of float values. The numpy array must be of shape
+                `(num_samples,)` for mono audio (`feature_size = 1`), or `(2, num_samples)` for stereo audio
+                (`feature_size = 2`).
+            padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `True`):
+                Select a strategy to pad the returned sequences (according to the model's padding side and padding
+                index) among:
+
+                - `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+                  sequence if provided).
+                - `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
+                  acceptable input length for the model if that argument is not provided.
+                - `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
+                  lengths).
+            truncation (`bool`, *optional*, defaults to `False`):
+                Activates truncation to cut input sequences longer than `max_length` to `max_length`.
+            max_length (`int`, *optional*):
+                Maximum length of the returned list and optionally padding length (see above).
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors instead of list of python integers. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return Numpy `np.ndarray` objects.
+            sampling_rate (`int`, *optional*):
+                The sampling rate at which the `audio` input was sampled. It is strongly recommended to pass
+                `sampling_rate` at the forward call to prevent silent errors.
+        """
+        if sampling_rate is not None:
+            if sampling_rate != self.sampling_rate:
+                raise ValueError(
+                    f"The model corresponding to this feature extractor: {self} was trained using a sampling rate of"
+                    f" {self.sampling_rate}. Please make sure that the provided audio input was sampled with"
+                    f" {self.sampling_rate} and not {sampling_rate}."
+                )
+        else:
+            logger.warning(
+                "It is strongly recommended to pass the `sampling_rate` argument to this function. "
+                "Failing to do so can result in silent errors that might be hard to debug."
+            )
+
+        if padding and truncation:
+            raise ValueError("Both padding and truncation were set. Make sure you only set one.")
+        elif padding is None:
+            # by default let's pad the inputs
+            padding = True
+
+        is_batched = bool(
+            isinstance(raw_audio, (list, tuple)) and (isinstance(raw_audio[0], (np.ndarray, tuple, list)))
+        )
+
+        if is_batched:
+            raw_audio = [np.asarray(audio, dtype=np.float32).T for audio in raw_audio]
+        elif not is_batched and not isinstance(raw_audio, np.ndarray):
+            raw_audio = np.asarray(raw_audio, dtype=np.float32)
+        elif isinstance(raw_audio, np.ndarray) and raw_audio.dtype is np.dtype(np.float64):
+            raw_audio = raw_audio.astype(np.float32)
+
+        # always return batch
+        if not is_batched:
+            raw_audio = [np.asarray(raw_audio).T]
+
+        # verify inputs are valid
+        for idx, example in enumerate(raw_audio):
+            if example.ndim > 2:
+                raise ValueError(f"Expected input shape (channels, length) but got shape {example.shape}")
+            if self.feature_size == 1 and example.ndim != 1:
+                raise ValueError(f"Expected mono audio but example has {example.shape[-1]} channels")
+            if self.feature_size == 2 and example.shape[-1] != 2:
+                raise ValueError(f"Expected stereo audio but example has {example.shape[-1]} channels")
+
+        padded_inputs = None
+        input_values = BatchFeature({"input_values": raw_audio})
+        if self.chunk_stride is not None and self.chunk_length is not None and max_length is None:
+            if truncation:
+                max_length = min(array.shape[0] for array in raw_audio)
+                nb_step = int(np.floor(max_length / self.chunk_stride))
+                max_length = (nb_step - 1) * self.chunk_stride + self.chunk_length
+            elif padding:
+                max_length = max(array.shape[0] for array in raw_audio)
+                nb_step = int(np.ceil(max_length / self.chunk_stride))
+                max_length = (nb_step - 1) * self.chunk_stride + self.chunk_length
+                padding = "max_length"
+            else:
+                padded_inputs = input_values
+
+        # normal padding on batch
+        if padded_inputs is None:
+            padded_inputs = self.pad(
+                input_values,
+                max_length=max_length,
+                truncation=truncation,
+                padding=padding,
+                return_attention_mask=padding,
+            )
+            if padding:
+                padded_inputs["padding_mask"] = padded_inputs.pop("attention_mask")
+
+        input_values = []
+        for example in padded_inputs.pop("input_values"):
+            if self.feature_size == 1:
+                example = example[..., None]
+            input_values.append(example.T)
+
+        padded_inputs["input_values"] = input_values
+        if return_tensors is not None:
+            padded_inputs = padded_inputs.convert_to_tensors(return_tensors)
+
+        return padded_inputs