Add new meta w2v2-conformer BERT-like model (#28165)

* first commit

* correct default value non causal

* update config and modeling code

* update converting checkpoint

* clean modeling and fix tests

* make style

* add new config parameters to docstring

* fix copied from statements

* Apply suggestions from code review
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* make position_embeddings_type docstrings clearer

* clean converting script

* remove function not used

* clean modeling file

* apply suggestion for test file + add convert script to not_doctested

* modify tests according to review - cleaner logic and more tests

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

* add checker of valid position embeddings type

* instantiate new layer norm layer with the right eps

* fix freeze_feature_encoder since it can be None in some cases

* add test same output in convert script

* restore wav2vec2conformer and add new model

* create processor and FE + clean

* add new model code

* fix convert script and set default config parameters

* correct model id paths

* make style

* make fix-copies and cleaning files

* fix copied from statements

* complete .md and fixe copies

* clean convert script argument defaults

* fix config parameters docstrings

* fix config docstring

* add copied from and enrich FE tests

* fix copied from and repo-consistency

* add autotokenizer

* make test input length shorter and change docstring code

* fix docstrings and copied from

* add add_adapter to ASR training example

* make testing of adapters more robust

* adapt to multi adapter layers

* refactor input_values->input_features and remove w2v2-bert feature extractor

* remove pretraining model

* remove depreciated features and useless lines

* add copied from and ignore statements to modeling tests

* remove pretraining model #2

* change import in convert script

* change default in convert script

* update readme and remove useless line

* Update tests/models/wav2vec2_bert/test_processor_wav2vec2_bert.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* refactor BERT to Bert for consistency

* remove useless ignore copy statement

* add persistent to buffer in rotary

* add eps in LayerNorm init and remove copied from

* add adapter activation parameters and add copied from statements

* Fix copied statements and add unitest.skip reasons

* add copied statement in test_processor

* refactor processor

* make style

* replace numpy random by torch rand

* remove expected output CTC

* improve converting script with processor class

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

* remove gumbel class

* remove tests related to previously deleted class

* Update src/transformers/models/wav2vec2_bert/configuration_wav2vec2_bert.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* correct typos

* remove uused parameters

* update processor to takes both text and audio

* update checkpoints

* update expected output and add ctc expected output

* add label_attention_mask

* replace pt with np in processor tests

* fix typo

* revert to behaviour with labels_attention_mask

---------
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

README.md

@@ -519,6 +519,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
+1. **[Wav2Vec2-BERT](https://huggingface.co/docs/transformers/main/model_doc/wav2vec2-bert)** (from Meta AI) released with the paper [Seamless: Multilingual Expressive and Streaming Speech Translation](https://ai.meta.com/research/publications/seamless-multilingual-expressive-and-streaming-speech-translation/) by the Seamless Communication team.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.
 1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/transformers/model_doc/wav2vec2_phoneme)** (from Facebook AI) released with the paper [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) by Qiantong Xu, Alexei Baevski, Michael Auli.
 1. **[WavLM](https://huggingface.co/docs/transformers/model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.

README_es.md

@@ -494,6 +494,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
+1. **[Wav2Vec2-BERT](https://huggingface.co/docs/transformers/main/model_doc/wav2vec2-bert)** (from Meta AI) released with the paper [Seamless: Multilingual Expressive and Streaming Speech Translation](https://ai.meta.com/research/publications/seamless-multilingual-expressive-and-streaming-speech-translation/) by the Seamless Communication team.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.
 1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/transformers/model_doc/wav2vec2_phoneme)** (from Facebook AI) released with the paper [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) by Qiantong Xu, Alexei Baevski, Michael Auli.
 1. **[WavLM](https://huggingface.co/docs/transformers/model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.

README_hd.md

@@ -468,6 +468,7 @@ conda install conda-forge::transformers
 1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (Kakao Enterprise से) Jaehyeon Kim, Jungil Kong, Juhee Son. द्वाराअनुसंधान पत्र [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) के साथ जारी किया गया
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (फेसबुक एआई से) साथ में पेपर [wav2vec 2.0: ए फ्रेमवर्क फॉर सेल्फ-सुपरवाइज्ड लर्निंग ऑफ स्पीच रिप्रेजेंटेशन](https://arxiv.org/abs/2006.11477) एलेक्सी बेवस्की, हेनरी झोउ, अब्देलरहमान मोहम्मद, माइकल औली द्वारा।
+1. **[Wav2Vec2-BERT](https://huggingface.co/docs/transformers/main/model_doc/wav2vec2-bert)** (from Meta AI) released with the paper [Seamless: Multilingual Expressive and Streaming Speech Translation](https://ai.meta.com/research/publications/seamless-multilingual-expressive-and-streaming-speech-translation/) by the Seamless Communication team.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (Facebook AI से) साथ वाला पेपर [FAIRSEQ S2T: FAIRSEQ के साथ फास्ट स्पीच-टू-टेक्स्ट मॉडलिंग ](https://arxiv.org/abs/2010.05171) चांगहान वांग, यूं तांग, जुताई मा, ऐनी वू, सरव्या पोपुरी, दिमित्रो ओखोनको, जुआन पिनो द्वारा पोस्ट किया गया।
 1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/transformers/model_doc/wav2vec2_phoneme)** (Facebook AI से) साथ वाला पेपर [सरल और प्रभावी जीरो-शॉट क्रॉस-लिंगुअल फोनेम रिकॉग्निशन](https://arxiv.org/abs/2109.11680) कियानटोंग जू, एलेक्सी बाएव्स्की, माइकल औली द्वारा।
 1. **[WavLM](https://huggingface.co/docs/transformers/model_doc/wavlm)** (माइक्रोसॉफ्ट रिसर्च से) पेपर के साथ जारी किया गया [WavLM: फुल स्टैक के लिए बड़े पैमाने पर स्व-पर्यवेक्षित पूर्व-प्रशिक्षण स्पीच प्रोसेसिंग](https://arxiv.org/abs/2110.13900) सानयुआन चेन, चेंगयी वांग, झेंगयांग चेन, यू वू, शुजी लियू, ज़ुओ चेन, जिन्यु ली, नाओयुकी कांडा, ताकुया योशियोका, ज़िओंग जिओ, जियान वू, लॉन्ग झोउ, शुओ रेन, यानमिन कियान, याओ कियान, जियान वू, माइकल ज़ेंग, फुरु वेई।

README_ja.md

@@ -528,6 +528,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (Kakao Enterprise から) Jaehyeon Kim, Jungil Kong, Juhee Son. から公開された研究論文 [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103)
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (Facebook AI から) Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli から公開された研究論文: [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477)
+1. **[Wav2Vec2-BERT](https://huggingface.co/docs/transformers/main/model_doc/wav2vec2-bert)** (from Meta AI) released with the paper [Seamless: Multilingual Expressive and Streaming Speech Translation](https://ai.meta.com/research/publications/seamless-multilingual-expressive-and-streaming-speech-translation/) by the Seamless Communication team.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (Facebook AI から) Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino から公開された研究論文: [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171)
 1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/transformers/model_doc/wav2vec2_phoneme)** (Facebook AI から) Qiantong Xu, Alexei Baevski, Michael Auli から公開された研究論文: [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680)
 1. **[WavLM](https://huggingface.co/docs/transformers/model_doc/wavlm)** (Microsoft Research から) Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei から公開された研究論文: [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900)

README_ko.md

@@ -443,6 +443,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (Kakao Enterprise 에서 제공)은 Jaehyeon Kim, Jungil Kong, Juhee Son.의 [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103)논문과 함께 발표했습니다.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (Facebook AI 에서) Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli 의 [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) 논문과 함께 발표했습니다.
+1. **[Wav2Vec2-BERT](https://huggingface.co/docs/transformers/main/model_doc/wav2vec2-bert)** (from Meta AI) released with the paper [Seamless: Multilingual Expressive and Streaming Speech Translation](https://ai.meta.com/research/publications/seamless-multilingual-expressive-and-streaming-speech-translation/) by the Seamless Communication team.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (Facebook AI 에서) Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino 의 [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) 논문과 함께 발표했습니다.
 1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/transformers/model_doc/wav2vec2_phoneme)** (Facebook AI 에서) Qiantong Xu, Alexei Baevski, Michael Auli 의 [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) 논문과 함께 발표했습니다.
 1. **[WavLM](https://huggingface.co/docs/transformers/model_doc/wavlm)** (Microsoft Research 에서) Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei 의 [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) 논문과 함께 발표했습니다.

README_zh-hans.md

@@ -467,6 +467,7 @@ conda install conda-forge::transformers
 1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (来自 Kakao Enterprise) 伴随论文 [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) 由 Jaehyeon Kim, Jungil Kong, Juhee Son 发布。
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (来自 Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) 由 Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (来自 Facebook AI) 伴随论文 [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) 由 Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli 发布。
+1. **[Wav2Vec2-BERT](https://huggingface.co/docs/transformers/main/model_doc/wav2vec2-bert)** (from Meta AI) released with the paper [Seamless: Multilingual Expressive and Streaming Speech Translation](https://ai.meta.com/research/publications/seamless-multilingual-expressive-and-streaming-speech-translation/) by the Seamless Communication team.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (来自 Facebook AI) 伴随论文 [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) 由 Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino 发布。
 1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/transformers/model_doc/wav2vec2_phoneme)** (来自 Facebook AI) 伴随论文 [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) 由 Qiantong Xu, Alexei Baevski, Michael Auli 发布。
 1. **[WavLM](https://huggingface.co/docs/transformers/model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.

README_zh-hant.md

@@ -479,6 +479,7 @@ conda install conda-forge::transformers
 1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
+1. **[Wav2Vec2-BERT](https://huggingface.co/docs/transformers/main/model_doc/wav2vec2-bert)** (from Meta AI) released with the paper [Seamless: Multilingual Expressive and Streaming Speech Translation](https://ai.meta.com/research/publications/seamless-multilingual-expressive-and-streaming-speech-translation/) by the Seamless Communication team.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.
 1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/transformers/model_doc/wav2vec2_phoneme)** (from Facebook AI) released with the paper [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) by Qiantong Xu, Alexei Baevski, Michael Auli.
 1. **[WavLM](https://huggingface.co/docs/transformers/model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.

docs/source/en/_toctree.yml

@@ -650,6 +650,8 @@
         title: VITS
       - local: model_doc/wav2vec2
         title: Wav2Vec2
+      - local: model_doc/wav2vec2-bert
+        title: Wav2Vec2-BERT
       - local: model_doc/wav2vec2-conformer
         title: Wav2Vec2-Conformer
       - local: model_doc/wav2vec2_phoneme

docs/source/en/index.md

@@ -296,6 +296,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                          [VITS](model_doc/vits)                          |       ✅        |         ❌         |      ❌      |
 |                         [ViViT](model_doc/vivit)                         |       ✅        |         ❌         |      ❌      |
 |                      [Wav2Vec2](model_doc/wav2vec2)                      |       ✅        |         ✅         |      ✅      |
+|                 [Wav2Vec2-BERT](model_doc/wav2vec2-bert)                 |       ✅        |         ❌         |      ❌      |
 |            [Wav2Vec2-Conformer](model_doc/wav2vec2-conformer)            |       ✅        |         ❌         |      ❌      |
 |              [Wav2Vec2Phoneme](model_doc/wav2vec2_phoneme)               |       ✅        |         ✅         |      ✅      |
 |                         [WavLM](model_doc/wavlm)                         |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/wav2vec2-bert.md

+<!--Copyright 2024 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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Wav2Vec2-BERT
+
+## Overview
+
+The Wav2Vec2-BERT model was proposed in [Seamless: Multilingual Expressive and Streaming Speech Translation](https://ai.meta.com/research/publications/seamless-multilingual-expressive-and-streaming-speech-translation/) by the Seamless Communication team from Meta AI.
+
+This model was pre-trained on 4.5M hours of unlabeled audio data covering more than 143 languages. It requires finetuning to be used for downstream tasks such as Automatic Speech Recognition (ASR), or Audio Classification.
+
+The official results of the model can be found in Section 3.2.1 of the paper.
+
+The abstract from the paper is the following:
+
+*Recent advancements in automatic speech translation have dramatically expanded language coverage, improved multimodal capabilities, and enabled a wide range of tasks and functionalities. That said, large-scale automatic speech translation systems today lack key features that help machine-mediated communication feel seamless when compared to human-to-human dialogue. In this work, we introduce a family of models that enable end-to-end expressive and multilingual translations in a streaming fashion. First, we contribute an improved version of the massively multilingual and multimodal SeamlessM4T model—SeamlessM4T v2. This newer model, incorporating an updated UnitY2 framework, was trained on more low-resource language data. The expanded version of SeamlessAlign adds 114,800 hours of automatically aligned data for a total of 76 languages. SeamlessM4T v2 provides the foundation on which our two newest models, SeamlessExpressive and SeamlessStreaming, are initiated. SeamlessExpressive enables translation that preserves vocal styles and prosody. Compared to previous efforts in expressive speech research, our work addresses certain underexplored aspects of prosody, such as speech rate and pauses, while also preserving the style of one’s voice. As for SeamlessStreaming, our model leverages the Efficient Monotonic Multihead Attention (EMMA) mechanism to generate low-latency target translations without waiting for complete source utterances. As the first of its kind, SeamlessStreaming enables simultaneous speech-to-speech/text translation for multiple source and target languages. To understand the performance of these models, we combined novel and modified versions of existing automatic metrics to evaluate prosody, latency, and robustness. For human evaluations, we adapted existing protocols tailored for measuring the most relevant attributes in the preservation of meaning, naturalness, and expressivity. To ensure that our models can be used safely and responsibly, we implemented the first known red-teaming effort for multimodal machine translation, a system for the detection and mitigation of added toxicity, a systematic evaluation of gender bias, and an inaudible localized watermarking mechanism designed to dampen the impact of deepfakes. Consequently, we bring major components from SeamlessExpressive and SeamlessStreaming together to form Seamless, the first publicly available system that unlocks expressive cross-lingual communication in real-time. In sum, Seamless gives us a pivotal look at the technical foundation needed to turn the Universal Speech Translator from a science fiction concept into a real-world technology. Finally, contributions in this work—including models, code, and a watermark detector—are publicly released and accessible at the link below.*
+
+This model was contributed by [ylacombe](https://huggingface.co/ylacombe). The original code can be found [here](https://github.com/facebookresearch/seamless_communication).
+
+## Usage tips
+
+- Wav2Vec2-BERT follows the same architecture as Wav2Vec2-Conformer, but employs a causal depthwise convolutional layer and uses as input a mel-spectrogram representation of the audio instead of the raw waveform.
+- Wav2Vec2-BERT can use either no relative position embeddings, Shaw-like position embeddings, Transformer-XL-like position embeddings, or
+  rotary position embeddings by setting the correct `config.position_embeddings_type`.
+- Wav2Vec2-BERT also introduces a Conformer-based adapter network instead of a simple convolutional network.
+
+## Resources
+
+<PipelineTag pipeline="automatic-speech-recognition"/>
+
+- [`Wav2Vec2BertForCTC`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/speech-recognition).
+- You can also adapt these notebooks on [how to finetune a speech recognition model in English](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/speech_recognition.ipynb), and [how to finetune a speech recognition model in any language](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/multi_lingual_speech_recognition.ipynb).
+
+<PipelineTag pipeline="audio-classification"/>
+
+- [`Wav2Vec2BertForSequenceClassification`] can be used by adapting this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/audio-classification).
+- See also: [Audio classification task guide](../tasks/audio_classification)
+
+
+## Wav2Vec2BertConfig
+
+[[autodoc]] Wav2Vec2BertConfig
+
+## Wav2Vec2BertProcessor
+
+[[autodoc]] Wav2Vec2BertProcessor
+    - __call__
+    - pad
+    - from_pretrained
+    - save_pretrained
+    - batch_decode
+    - decode
+
+## Wav2Vec2BertModel
+
+[[autodoc]] Wav2Vec2BertModel
+    - forward
+
+## Wav2Vec2BertForCTC
+
+[[autodoc]] Wav2Vec2BertForCTC
+    - forward
+
+## Wav2Vec2BertForSequenceClassification
+
+[[autodoc]] Wav2Vec2BertForSequenceClassification
+    - forward
+
+## Wav2Vec2BertForAudioFrameClassification
+
+[[autodoc]] Wav2Vec2BertForAudioFrameClassification
+    - forward
+
+## Wav2Vec2BertForXVector
+
+[[autodoc]] Wav2Vec2BertForXVector
+    - forward

docs/source/en/tasks/asr.md

@@ -32,7 +32,7 @@ The task illustrated in this tutorial is supported by the following model archit
 
 <!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
 
-[Data2VecAudio](../model_doc/data2vec-audio), [Hubert](../model_doc/hubert), [M-CTC-T](../model_doc/mctct), [SEW](../model_doc/sew), [SEW-D](../model_doc/sew-d), [UniSpeech](../model_doc/unispeech), [UniSpeechSat](../model_doc/unispeech-sat), [Wav2Vec2](../model_doc/wav2vec2), [Wav2Vec2-Conformer](../model_doc/wav2vec2-conformer), [WavLM](../model_doc/wavlm)
+[Data2VecAudio](../model_doc/data2vec-audio), [Hubert](../model_doc/hubert), [M-CTC-T](../model_doc/mctct), [SEW](../model_doc/sew), [SEW-D](../model_doc/sew-d), [UniSpeech](../model_doc/unispeech), [UniSpeechSat](../model_doc/unispeech-sat), [Wav2Vec2](../model_doc/wav2vec2), [Wav2Vec2-BERT](../model_doc/wav2vec2-bert), [Wav2Vec2-Conformer](../model_doc/wav2vec2-conformer), [WavLM](../model_doc/wavlm)
 
 <!--End of the generated tip-->
 

docs/source/en/tasks/audio_classification.md

@@ -32,7 +32,7 @@ The task illustrated in this tutorial is supported by the following model archit
 
 <!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
 
-[Audio Spectrogram Transformer](../model_doc/audio-spectrogram-transformer), [Data2VecAudio](../model_doc/data2vec-audio), [Hubert](../model_doc/hubert), [SEW](../model_doc/sew), [SEW-D](../model_doc/sew-d), [UniSpeech](../model_doc/unispeech), [UniSpeechSat](../model_doc/unispeech-sat), [Wav2Vec2](../model_doc/wav2vec2), [Wav2Vec2-Conformer](../model_doc/wav2vec2-conformer), [WavLM](../model_doc/wavlm), [Whisper](../model_doc/whisper)
+[Audio Spectrogram Transformer](../model_doc/audio-spectrogram-transformer), [Data2VecAudio](../model_doc/data2vec-audio), [Hubert](../model_doc/hubert), [SEW](../model_doc/sew), [SEW-D](../model_doc/sew-d), [UniSpeech](../model_doc/unispeech), [UniSpeechSat](../model_doc/unispeech-sat), [Wav2Vec2](../model_doc/wav2vec2), [Wav2Vec2-BERT](../model_doc/wav2vec2-bert), [Wav2Vec2-Conformer](../model_doc/wav2vec2-conformer), [WavLM](../model_doc/wavlm), [Whisper](../model_doc/whisper)
 
 <!--End of the generated tip-->
 

examples/pytorch/speech-recognition/run_speech_recognition_ctc.py

@@ -132,6 +132,13 @@ class ModelArguments:
     ctc_loss_reduction: Optional[str] = field(
         default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
     )
+    add_adapter: Optional[bool] = field(
+        default=False,
+        metadata={
+            "help": "Whether a convolutional attention network should be stacked on top of the Wav2Vec2BERT Encoder. Can be very"
+            "useful to downsample the output length."
+        },
+    )
 
 
 @dataclass
@@ -602,6 +609,7 @@ def main():
             "pad_token_id": tokenizer.pad_token_id,
             "vocab_size": len(tokenizer),
             "activation_dropout": model_args.activation_dropout,
+            "add_adapter": model_args.add_adapter,
         }
     )
 

src/transformers/__init__.py

@@ -914,6 +914,11 @@ _import_structure = {
         "Wav2Vec2Processor",
         "Wav2Vec2Tokenizer",
     ],
+    "models.wav2vec2_bert": [
+        "WAV2VEC2_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "Wav2Vec2BertConfig",
+        "Wav2Vec2BertProcessor",
+    ],
     "models.wav2vec2_conformer": [
         "WAV2VEC2_CONFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP",
         "Wav2Vec2ConformerConfig",
@@ -3515,6 +3520,17 @@ else:
             "Wav2Vec2PreTrainedModel",
         ]
     )
+    _import_structure["models.wav2vec2_bert"].extend(
+        [
+            "WAV2VEC2_BERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "Wav2Vec2BertForAudioFrameClassification",
+            "Wav2Vec2BertForCTC",
+            "Wav2Vec2BertForSequenceClassification",
+            "Wav2Vec2BertForXVector",
+            "Wav2Vec2BertModel",
+            "Wav2Vec2BertPreTrainedModel",
+        ]
+    )
     _import_structure["models.wav2vec2_conformer"].extend(
         [
             "WAV2VEC2_CONFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -5617,6 +5633,11 @@ if TYPE_CHECKING:
         Wav2Vec2Processor,
         Wav2Vec2Tokenizer,
     )
+    from .models.wav2vec2_bert import (
+        WAV2VEC2_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        Wav2Vec2BertConfig,
+        Wav2Vec2BertProcessor,
+    )
     from .models.wav2vec2_conformer import (
         WAV2VEC2_CONFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP,
         Wav2Vec2ConformerConfig,
@@ -7821,6 +7842,15 @@ if TYPE_CHECKING:
             Wav2Vec2Model,
             Wav2Vec2PreTrainedModel,
         )
+        from .models.wav2vec2_bert import (
+            WAV2VEC2_BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            Wav2Vec2BertForAudioFrameClassification,
+            Wav2Vec2BertForCTC,
+            Wav2Vec2BertForSequenceClassification,
+            Wav2Vec2BertForXVector,
+            Wav2Vec2BertModel,
+            Wav2Vec2BertPreTrainedModel,
+        )
         from .models.wav2vec2_conformer import (
             WAV2VEC2_CONFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             Wav2Vec2ConformerForAudioFrameClassification,

src/transformers/models/__init__.py

@@ -235,6 +235,7 @@ from . import (
     vits,
     vivit,
     wav2vec2,
+    wav2vec2_bert,
     wav2vec2_conformer,
     wav2vec2_phoneme,
     wav2vec2_with_lm,

src/transformers/models/auto/configuration_auto.py

@@ -246,6 +246,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("vits", "VitsConfig"),
         ("vivit", "VivitConfig"),
         ("wav2vec2", "Wav2Vec2Config"),
+        ("wav2vec2-bert", "Wav2Vec2BertConfig"),
         ("wav2vec2-conformer", "Wav2Vec2ConformerConfig"),
         ("wavlm", "WavLMConfig"),
         ("whisper", "WhisperConfig"),
@@ -459,6 +460,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("vits", "VITS_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("vivit", "VIVIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("wav2vec2", "WAV_2_VEC_2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("wav2vec2-bert", "WAV2VEC2_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("wav2vec2-conformer", "WAV2VEC2_CONFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("whisper", "WHISPER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("xclip", "XCLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -718,6 +720,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("vits", "VITS"),
         ("vivit", "ViViT"),
         ("wav2vec2", "Wav2Vec2"),
+        ("wav2vec2-bert", "Wav2Vec2-BERT"),
         ("wav2vec2-conformer", "Wav2Vec2-Conformer"),
         ("wav2vec2_phoneme", "Wav2Vec2Phoneme"),
         ("wavlm", "WavLM"),

src/transformers/models/auto/feature_extraction_auto.py

@@ -100,6 +100,7 @@ FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict(
         ("vit_mae", "ViTFeatureExtractor"),
         ("vit_msn", "ViTFeatureExtractor"),
         ("wav2vec2", "Wav2Vec2FeatureExtractor"),
+        ("wav2vec2-bert", "Wav2Vec2FeatureExtractor"),
         ("wav2vec2-conformer", "Wav2Vec2FeatureExtractor"),
         ("wavlm", "Wav2Vec2FeatureExtractor"),
         ("whisper", "WhisperFeatureExtractor"),

src/transformers/models/auto/modeling_auto.py

@@ -232,6 +232,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("vits", "VitsModel"),
         ("vivit", "VivitModel"),
         ("wav2vec2", "Wav2Vec2Model"),
+        ("wav2vec2-bert", "Wav2Vec2BertModel"),
         ("wav2vec2-conformer", "Wav2Vec2ConformerModel"),
         ("wavlm", "WavLMModel"),
         ("whisper", "WhisperModel"),
@@ -1034,6 +1035,7 @@ MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("unispeech", "UniSpeechForSequenceClassification"),
         ("unispeech-sat", "UniSpeechSatForSequenceClassification"),
         ("wav2vec2", "Wav2Vec2ForSequenceClassification"),
+        ("wav2vec2-bert", "Wav2Vec2BertForSequenceClassification"),
         ("wav2vec2-conformer", "Wav2Vec2ConformerForSequenceClassification"),
         ("wavlm", "WavLMForSequenceClassification"),
         ("whisper", "WhisperForAudioClassification"),
@@ -1051,6 +1053,7 @@ MODEL_FOR_CTC_MAPPING_NAMES = OrderedDict(
         ("unispeech", "UniSpeechForCTC"),
         ("unispeech-sat", "UniSpeechSatForCTC"),
         ("wav2vec2", "Wav2Vec2ForCTC"),
+        ("wav2vec2-bert", "Wav2Vec2BertForCTC"),
         ("wav2vec2-conformer", "Wav2Vec2ConformerForCTC"),
         ("wavlm", "WavLMForCTC"),
     ]
@@ -1062,6 +1065,7 @@ MODEL_FOR_AUDIO_FRAME_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("data2vec-audio", "Data2VecAudioForAudioFrameClassification"),
         ("unispeech-sat", "UniSpeechSatForAudioFrameClassification"),
         ("wav2vec2", "Wav2Vec2ForAudioFrameClassification"),
+        ("wav2vec2-bert", "Wav2Vec2BertForAudioFrameClassification"),
         ("wav2vec2-conformer", "Wav2Vec2ConformerForAudioFrameClassification"),
         ("wavlm", "WavLMForAudioFrameClassification"),
     ]
@@ -1073,6 +1077,7 @@ MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES = OrderedDict(
         ("data2vec-audio", "Data2VecAudioForXVector"),
         ("unispeech-sat", "UniSpeechSatForXVector"),
         ("wav2vec2", "Wav2Vec2ForXVector"),
+        ("wav2vec2-bert", "Wav2Vec2BertForXVector"),
         ("wav2vec2-conformer", "Wav2Vec2ConformerForXVector"),
         ("wavlm", "WavLMForXVector"),
     ]

src/transformers/models/auto/processing_auto.py

@@ -91,6 +91,7 @@ PROCESSOR_MAPPING_NAMES = OrderedDict(
         ("vipllava", "LlavaProcessor"),
         ("vision-text-dual-encoder", "VisionTextDualEncoderProcessor"),
         ("wav2vec2", "Wav2Vec2Processor"),
+        ("wav2vec2-bert", "Wav2Vec2Processor"),
         ("wav2vec2-conformer", "Wav2Vec2Processor"),
         ("wavlm", "Wav2Vec2Processor"),
         ("whisper", "WhisperProcessor"),

src/transformers/models/auto/tokenization_auto.py

@@ -418,6 +418,7 @@ else:
             ("visual_bert", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
             ("vits", ("VitsTokenizer", None)),
             ("wav2vec2", ("Wav2Vec2CTCTokenizer", None)),
+            ("wav2vec2-bert", ("Wav2Vec2CTCTokenizer", None)),
             ("wav2vec2-conformer", ("Wav2Vec2CTCTokenizer", None)),
             ("wav2vec2_phoneme", ("Wav2Vec2PhonemeCTCTokenizer", None)),
             ("whisper", ("WhisperTokenizer", "WhisperTokenizerFast" if is_tokenizers_available() else None)),