Add SeamlessM4T v2 (#27779)

* add working convertion script

* first non-working version of modeling code

* update modeling code (working)

* make style

* make fix-copies

* add config docstrings

* add config to ignore docstrings formatage due to unconventional markdown

* fix copies

* fix generation num_return_sequences

* enrich docs

* add and fix tests beside integration tests

* update integration tests

* update repo id

* add tie weights and make style

* correct naming in .md

* fix imports and so on

* correct docstrings

* fix fp16 speech forward

* fix speechencoder attention

* make style

* fix copied from

* rename SeamlessM4Tv2-v2 to SeamlessM4Tv2

* Apply suggestions on configuration
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* remove useless public models

* fix private models + better naming for T2U models

* clean speech encoder relative position embeddings

* refactor chunk attention

* add docstrings to chunk attention method

* improve naming and docstrings

* rename some attention variables + add temperature sampling in T2U model

* rename DOCSTRINGS variable names

* make style + remove 2 useless config parameters

* enrich model card

* remove any attention_head reference + fix temperature in T2U

* new fmt and make style

* Apply suggestions from code review
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* rename spkr_id->speaker_id and change docstrings of get_char_input_ids

* simplify v2attention

* make style

* Update seamless_m4t_v2.md

* update code and tests with last update

* update repo ids

* fill article name, abstract andauthors

* update not_doctested and slow_doc tests

---------
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

README.md

@@ -466,6 +466,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer)** (from ZhuiyiTechnology), released together with the paper [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/abs/2104.09864) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
 1. **[RWKV](https://huggingface.co/docs/transformers/model_doc/rwkv)** (from Bo Peng), released on [this repo](https://github.com/BlinkDL/RWKV-LM) by Bo Peng.
 1. **[SeamlessM4T](https://huggingface.co/docs/transformers/model_doc/seamless_m4t)** (from Meta AI) released with the paper [SeamlessM4T — Massively Multilingual & Multimodal Machine Translation](https://dl.fbaipublicfiles.com/seamless/seamless_m4t_paper.pdf) by the Seamless Communication team.
+1. **[SeamlessM4Tv2](https://huggingface.co/docs/transformers/main/model_doc/seamless_m4t_v2)** (from <FILL INSTITUTION>) released with the paper [<FILL PAPER TITLE>](<FILL ARKIV LINK>) by <FILL AUTHORS>.
 1. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo.
 1. **[Segment Anything](https://huggingface.co/docs/transformers/model_doc/sam)** (from Meta AI) released with the paper [Segment Anything](https://arxiv.org/pdf/2304.02643v1.pdf) by Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick.
 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.

README_es.md

@@ -441,6 +441,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer)** (from ZhuiyiTechnology), released together with the paper [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/abs/2104.09864) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
 1. **[RWKV](https://huggingface.co/docs/transformers/model_doc/rwkv)** (from Bo Peng) released with the paper [this repo](https://github.com/BlinkDL/RWKV-LM) by Bo Peng.
 1. **[SeamlessM4T](https://huggingface.co/docs/transformers/model_doc/seamless_m4t)** (from Meta AI) released with the paper [SeamlessM4T — Massively Multilingual & Multimodal Machine Translation](https://dl.fbaipublicfiles.com/seamless/seamless_m4t_paper.pdf) by the Seamless Communication team.
+1. **[SeamlessM4Tv2](https://huggingface.co/docs/transformers/main/model_doc/seamless_m4t_v2)** (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. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo.
 1. **[Segment Anything](https://huggingface.co/docs/transformers/model_doc/sam)** (from Meta AI) released with the paper [Segment Anything](https://arxiv.org/pdf/2304.02643v1.pdf) by Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick.
 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.

README_hd.md

@@ -415,6 +415,7 @@ conda install -c huggingface transformers
 1. **[RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer)** (झुईई टेक्नोलॉजी से), साथ में पेपर [रोफॉर्मर: रोटरी पोजिशन एंबेडिंग के साथ एन्हांस्ड ट्रांसफॉर्मर] (https://arxiv.org/pdf/2104.09864v1.pdf) जियानलिन सु और यू लू और शेंगफेंग पैन और बो वेन और युनफेंग लियू द्वारा प्रकाशित।
 1. **[RWKV](https://huggingface.co/docs/transformers/model_doc/rwkv)** (Bo Peng से) Bo Peng. द्वाराअनुसंधान पत्र [this repo](https://github.com/BlinkDL/RWKV-LM) के साथ जारी किया गया
 1. **[SeamlessM4T](https://huggingface.co/docs/transformers/model_doc/seamless_m4t)** (from Meta AI) released with the paper [SeamlessM4T — Massively Multilingual & Multimodal Machine Translation](https://dl.fbaipublicfiles.com/seamless/seamless_m4t_paper.pdf) by the Seamless Communication team.
+1. **[SeamlessM4Tv2](https://huggingface.co/docs/transformers/main/model_doc/seamless_m4t_v2)** (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. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo.
 1. **[Segment Anything](https://huggingface.co/docs/transformers/model_doc/sam)** (Meta AI से) Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick. द्वाराअनुसंधान पत्र [Segment Anything](https://arxiv.org/pdf/2304.02643v1.pdf) के साथ जारी किया गया
 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (ASAPP से) साथ देने वाला पेपर [भाषण पहचान के लिए अनसुपरवाइज्ड प्री-ट्रेनिंग में परफॉर्मेंस-एफिशिएंसी ट्रेड-ऑफ्स](https ://arxiv.org/abs/2109.06870) फेलिक्स वू, क्वांगयुन किम, जिंग पैन, क्यू हान, किलियन क्यू. वेनबर्गर, योव आर्टज़ी द्वारा।

README_ja.md

@@ -475,6 +475,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer)** (ZhuiyiTechnology から), Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu から公開された研究論文: [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/abs/2104.09864)
 1. **[RWKV](https://huggingface.co/docs/transformers/model_doc/rwkv)** (Bo Peng から) Bo Peng. から公開された研究論文 [this repo](https://github.com/BlinkDL/RWKV-LM)
 1. **[SeamlessM4T](https://huggingface.co/docs/transformers/model_doc/seamless_m4t)** (from Meta AI) released with the paper [SeamlessM4T — Massively Multilingual & Multimodal Machine Translation](https://dl.fbaipublicfiles.com/seamless/seamless_m4t_paper.pdf) by the Seamless Communication team.
+1. **[SeamlessM4Tv2](https://huggingface.co/docs/transformers/main/model_doc/seamless_m4t_v2)** (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. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (NVIDIA から) Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo から公開された研究論文: [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203)
 1. **[Segment Anything](https://huggingface.co/docs/transformers/model_doc/sam)** (Meta AI から) Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick. から公開された研究論文 [Segment Anything](https://arxiv.org/pdf/2304.02643v1.pdf)
 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (ASAPP から) Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi から公開された研究論文: [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870)

README_ko.md

@@ -390,6 +390,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer)** (ZhuiyiTechnology 에서) Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu 의 a [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/pdf/2104.09864v1.pdf) 논문과 함께 발표했습니다.
 1. **[RWKV](https://huggingface.co/docs/transformers/model_doc/rwkv)** (Bo Peng 에서 제공)은 Bo Peng.의 [this repo](https://github.com/BlinkDL/RWKV-LM)논문과 함께 발표했습니다.
 1. **[SeamlessM4T](https://huggingface.co/docs/transformers/model_doc/seamless_m4t)** (from Meta AI) released with the paper [SeamlessM4T — Massively Multilingual & Multimodal Machine Translation](https://dl.fbaipublicfiles.com/seamless/seamless_m4t_paper.pdf) by the Seamless Communication team.
+1. **[SeamlessM4Tv2](https://huggingface.co/docs/transformers/main/model_doc/seamless_m4t_v2)** (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. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (NVIDIA 에서) Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo 의 [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) 논문과 함께 발표했습니다.
 1. **[Segment Anything](https://huggingface.co/docs/transformers/model_doc/sam)** (Meta AI 에서 제공)은 Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick.의 [Segment Anything](https://arxiv.org/pdf/2304.02643v1.pdf)논문과 함께 발표했습니다.
 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (ASAPP 에서) Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi 의 [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) 논문과 함께 발표했습니다.

README_zh-hans.md

@@ -414,6 +414,7 @@ conda install -c huggingface transformers
 1. **[RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer)** (来自 ZhuiyiTechnology), 伴随论文 [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/pdf/2104.09864v1.pdf) 由 Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu 发布。
 1. **[RWKV](https://huggingface.co/docs/transformers/model_doc/rwkv)** (来自 Bo Peng) 伴随论文 [this repo](https://github.com/BlinkDL/RWKV-LM) 由 Bo Peng 发布。
 1. **[SeamlessM4T](https://huggingface.co/docs/transformers/model_doc/seamless_m4t)** (from Meta AI) released with the paper [SeamlessM4T — Massively Multilingual & Multimodal Machine Translation](https://dl.fbaipublicfiles.com/seamless/seamless_m4t_paper.pdf) by the Seamless Communication team.
+1. **[SeamlessM4Tv2](https://huggingface.co/docs/transformers/main/model_doc/seamless_m4t_v2)** (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. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (来自 NVIDIA) 伴随论文 [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) 由 Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo 发布。
 1. **[Segment Anything](https://huggingface.co/docs/transformers/model_doc/sam)** (来自 Meta AI) 伴随论文 [Segment Anything](https://arxiv.org/pdf/2304.02643v1.pdf) 由 Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick 发布。
 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (来自 ASAPP) 伴随论文 [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) 由 Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi 发布。

README_zh-hant.md

@@ -426,6 +426,7 @@ conda install -c huggingface transformers
 1. **[RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer)** (from ZhuiyiTechnology), released together with the paper a [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/pdf/2104.09864v1.pdf) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
 1. **[RWKV](https://huggingface.co/docs/transformers/model_doc/rwkv)** (from Bo Peng) released with the paper [this repo](https://github.com/BlinkDL/RWKV-LM) by Bo Peng.
 1. **[SeamlessM4T](https://huggingface.co/docs/transformers/model_doc/seamless_m4t)** (from Meta AI) released with the paper [SeamlessM4T — Massively Multilingual & Multimodal Machine Translation](https://dl.fbaipublicfiles.com/seamless/seamless_m4t_paper.pdf) by the Seamless Communication team.
+1. **[SeamlessM4Tv2](https://huggingface.co/docs/transformers/main/model_doc/seamless_m4t_v2)** (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. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo.
 1. **[Segment Anything](https://huggingface.co/docs/transformers/model_doc/sam)** (from Meta AI) released with the paper [Segment Anything](https://arxiv.org/pdf/2304.02643v1.pdf) by Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick.
 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.

docs/source/en/_toctree.yml

@@ -618,6 +618,8 @@
         title: Pop2Piano
       - local: model_doc/seamless_m4t
         title: Seamless-M4T
+      - local: model_doc/seamless_m4t_v2
+        title: SeamlessM4T-v2
       - local: model_doc/sew
         title: SEW
       - local: model_doc/sew-d

docs/source/en/index.md

@@ -242,6 +242,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                          [RWKV](model_doc/rwkv)                          |       ✅        |         ❌         |      ❌      |
 |                           [SAM](model_doc/sam)                           |       ✅        |         ✅         |      ❌      |
 |                  [SeamlessM4T](model_doc/seamless_m4t)                   |       ✅        |         ❌         |      ❌      |
+|                [SeamlessM4Tv2](model_doc/seamless_m4t_v2)                |       ✅        |         ❌         |      ❌      |
 |                     [SegFormer](model_doc/segformer)                     |       ✅        |         ✅         |      ❌      |
 |                           [SEW](model_doc/sew)                           |       ✅        |         ❌         |      ❌      |
 |                         [SEW-D](model_doc/sew-d)                         |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/seamless_m4t_v2.md

+<!--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.
+-->
+
+# SeamlessM4T-v2
+
+## Overview
+
+The SeamlessM4T-v2 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.
+
+SeamlessM4T-v2 is a collection of models designed to provide high quality translation, allowing people from different linguistic communities to communicate effortlessly through speech and text. It is an improvement on the [previous version](./seamless_m4t.md). For more details on the differences between v1 and v2, refer to section [Difference with SeamlessM4T-v1](#difference-with-seamlessm4t-v1).
+
+SeamlessM4T-v2 enables multiple tasks without relying on separate models:
+
+- Speech-to-speech translation (S2ST)
+- Speech-to-text translation (S2TT)
+- Text-to-speech translation (T2ST)
+- Text-to-text translation (T2TT)
+- Automatic speech recognition (ASR)
+
+[`SeamlessM4Tv2Model`] can perform all the above tasks, but each task also has its own dedicated sub-model.
+
+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.*
+
+## Usage
+
+In the following example, we'll load an Arabic audio sample and an English text sample and convert them into Russian speech and French text.
+
+First, load the processor and a checkpoint of the model:
+
+```python
+>>> from transformers import AutoProcessor, SeamlessM4Tv2Model
+
+>>> processor = AutoProcessor.from_pretrained("facebook/seamless-m4t-v2-large")
+>>> model = SeamlessM4Tv2Model.from_pretrained("facebook/seamless-m4t-v2-large")
+```
+
+You can seamlessly use this model on text or on audio, to generated either translated text or translated audio.
+
+Here is how to use the processor to process text and audio:
+
+```python
+>>> # let's load an audio sample from an Arabic speech corpus
+>>> from datasets import load_dataset
+>>> dataset = load_dataset("arabic_speech_corpus", split="test", streaming=True)
+>>> audio_sample = next(iter(dataset))["audio"]
+
+>>> # now, process it
+>>> audio_inputs = processor(audios=audio_sample["array"], return_tensors="pt")
+
+>>> # now, process some English text as well
+>>> text_inputs = processor(text = "Hello, my dog is cute", src_lang="eng", return_tensors="pt")
+```
+
+
+### Speech
+
+[`SeamlessM4Tv2Model`] can *seamlessly* generate text or speech with few or no changes. Let's target Russian voice translation:
+
+```python
+>>> audio_array_from_text = model.generate(**text_inputs, tgt_lang="rus")[0].cpu().numpy().squeeze()
+>>> audio_array_from_audio = model.generate(**audio_inputs, tgt_lang="rus")[0].cpu().numpy().squeeze()
+```
+
+With basically the same code, I've translated English text and Arabic speech to Russian speech samples.
+
+### Text
+
+Similarly, you can generate translated text from audio files or from text with the same model. You only have to pass `generate_speech=False` to [`SeamlessM4Tv2Model.generate`].
+This time, let's translate to French.
+
+```python 
+>>> # from audio
+>>> output_tokens = model.generate(**audio_inputs, tgt_lang="fra", generate_speech=False)
+>>> translated_text_from_audio = processor.decode(output_tokens[0].tolist()[0], skip_special_tokens=True)
+
+>>> # from text
+>>> output_tokens = model.generate(**text_inputs, tgt_lang="fra", generate_speech=False)
+>>> translated_text_from_text = processor.decode(output_tokens[0].tolist()[0], skip_special_tokens=True)
+```
+
+### Tips
+
+
+#### 1. Use dedicated models
+
+[`SeamlessM4Tv2Model`] is transformers top level model to generate speech and text, but you can also use dedicated models that perform the task without additional components, thus reducing the memory footprint.
+For example, you can replace the audio-to-audio generation snippet with the model dedicated to the S2ST task, the rest is exactly the same code: 
+
+```python
+>>> from transformers import SeamlessM4Tv2ForSpeechToSpeech
+>>> model = SeamlessM4Tv2ForSpeechToSpeech.from_pretrained("facebook/seamless-m4t-v2-large")
+```
+
+Or you can replace the text-to-text generation snippet with the model dedicated to the T2TT task, you only have to remove `generate_speech=False`.
+
+```python
+>>> from transformers import SeamlessM4Tv2ForTextToText
+>>> model = SeamlessM4Tv2ForTextToText.from_pretrained("facebook/seamless-m4t-v2-large")
+```
+
+Feel free to try out [`SeamlessM4Tv2ForSpeechToText`] and [`SeamlessM4Tv2ForTextToSpeech`] as well.
+
+#### 2. Change the speaker identity
+
+You have the possibility to change the speaker used for speech synthesis with the `speaker_id` argument. Some `speaker_id` works better than other for some languages!
+
+#### 3. Change the generation strategy
+
+You can use different [generation strategies](../generation_strategies) for text generation, e.g `.generate(input_ids=input_ids, text_num_beams=4, text_do_sample=True)` which will perform multinomial beam-search decoding on the text model. Note that speech generation only supports greedy - by default - or multinomial sampling, which can be used with e.g. `.generate(..., speech_do_sample=True, speech_temperature=0.6)`.
+
+#### 4. Generate speech and text at the same time
+
+Use `return_intermediate_token_ids=True` with [`SeamlessM4Tv2Model`] to return both speech and text !
+
+## Model architecture
+
+SeamlessM4T-v2 features a versatile architecture that smoothly handles the sequential generation of text and speech. This setup comprises two sequence-to-sequence (seq2seq) models. The first model translates the input modality into translated text, while the second model generates speech tokens, known as "unit tokens," from the translated text.
+
+Each modality has its own dedicated encoder with a unique architecture. Additionally, for speech output, a vocoder inspired by the [HiFi-GAN](https://arxiv.org/abs/2010.05646) architecture is placed on top of the second seq2seq model.
+
+### Difference with SeamlessM4T-v1
+
+The architecture of this new version differs from the first in a few aspects:
+
+#### Improvements on the second-pass model
+
+The second seq2seq model, named text-to-unit model, is now non-auto regressive, meaning that it computes units in a **single forward pass**. This achievement is made possible by:
+- the use of **character-level embeddings**, meaning that each character of the predicted translated text has its own embeddings, which are then used to predict the unit tokens.
+- the use of an intermediate duration predictor, that predicts speech duration at the **character-level** on the predicted translated text.
+- the use of a new text-to-unit decoder mixing convolutions and self-attention to handle longer context.
+
+#### Difference in the speech encoder
+
+The speech encoder, which is used during the first-pass generation process to predict the translated text, differs mainly from the previous speech encoder through these mechanisms:
+- the use of chunked attention mask to prevent attention across chunks, ensuring that each position attends only to positions within its own chunk and a fixed number of previous chunks.
+- the use of relative position embeddings which only considers distance between sequence elements rather than absolute positions. Please refer to [Self-Attentionwith Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155) for more details.
+- the use of a causal depth-wise convolution instead of a non-causal one.
+
+### Generation process
+
+Here's how the generation process works:
+
+- Input text or speech is processed through its specific encoder.
+- A decoder creates text tokens in the desired language.
+- If speech generation is required, the second seq2seq model, generates unit tokens in an non auto-regressive way.
+- These unit tokens are then passed through the final vocoder to produce the actual speech.
+
+
+This model was contributed by [ylacombe](https://huggingface.co/ylacombe). The original code can be found [here](https://github.com/facebookresearch/seamless_communication).
+
+## SeamlessM4Tv2Model
+
+[[autodoc]] SeamlessM4Tv2Model
+    - generate
+
+
+## SeamlessM4Tv2ForTextToSpeech
+
+[[autodoc]] SeamlessM4Tv2ForTextToSpeech
+    - generate
+
+
+## SeamlessM4Tv2ForSpeechToSpeech
+
+[[autodoc]] SeamlessM4Tv2ForSpeechToSpeech
+    - generate
+
+
+## SeamlessM4Tv2ForTextToText
+
+[[autodoc]] transformers.SeamlessM4Tv2ForTextToText
+    - forward
+    - generate
+
+## SeamlessM4Tv2ForSpeechToText
+
+[[autodoc]] transformers.SeamlessM4Tv2ForSpeechToText
+    - forward
+    - generate
+
+## SeamlessM4Tv2Config
+
+[[autodoc]] SeamlessM4Tv2Config

docs/source/en/tasks/summarization.md

@@ -35,7 +35,7 @@ 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!-->
 
-[BART](../model_doc/bart), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [Encoder decoder](../model_doc/encoder-decoder), [FairSeq Machine-Translation](../model_doc/fsmt), [GPTSAN-japanese](../model_doc/gptsan-japanese), [LED](../model_doc/led), [LongT5](../model_doc/longt5), [M2M100](../model_doc/m2m_100), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [NLLB](../model_doc/nllb), [NLLB-MOE](../model_doc/nllb-moe), [Pegasus](../model_doc/pegasus), [PEGASUS-X](../model_doc/pegasus_x), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [SeamlessM4T](../model_doc/seamless_m4t), [SwitchTransformers](../model_doc/switch_transformers), [T5](../model_doc/t5), [UMT5](../model_doc/umt5), [XLM-ProphetNet](../model_doc/xlm-prophetnet)
+[BART](../model_doc/bart), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [Encoder decoder](../model_doc/encoder-decoder), [FairSeq Machine-Translation](../model_doc/fsmt), [GPTSAN-japanese](../model_doc/gptsan-japanese), [LED](../model_doc/led), [LongT5](../model_doc/longt5), [M2M100](../model_doc/m2m_100), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [NLLB](../model_doc/nllb), [NLLB-MOE](../model_doc/nllb-moe), [Pegasus](../model_doc/pegasus), [PEGASUS-X](../model_doc/pegasus_x), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [SeamlessM4T](../model_doc/seamless_m4t), [SeamlessM4Tv2](../model_doc/seamless_m4t_v2), [SwitchTransformers](../model_doc/switch_transformers), [T5](../model_doc/t5), [UMT5](../model_doc/umt5), [XLM-ProphetNet](../model_doc/xlm-prophetnet)
 
 <!--End of the generated tip-->
 

docs/source/en/tasks/translation.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!-->
 
-[BART](../model_doc/bart), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [Encoder decoder](../model_doc/encoder-decoder), [FairSeq Machine-Translation](../model_doc/fsmt), [GPTSAN-japanese](../model_doc/gptsan-japanese), [LED](../model_doc/led), [LongT5](../model_doc/longt5), [M2M100](../model_doc/m2m_100), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [NLLB](../model_doc/nllb), [NLLB-MOE](../model_doc/nllb-moe), [Pegasus](../model_doc/pegasus), [PEGASUS-X](../model_doc/pegasus_x), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [SeamlessM4T](../model_doc/seamless_m4t), [SwitchTransformers](../model_doc/switch_transformers), [T5](../model_doc/t5), [UMT5](../model_doc/umt5), [XLM-ProphetNet](../model_doc/xlm-prophetnet)
+[BART](../model_doc/bart), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [Encoder decoder](../model_doc/encoder-decoder), [FairSeq Machine-Translation](../model_doc/fsmt), [GPTSAN-japanese](../model_doc/gptsan-japanese), [LED](../model_doc/led), [LongT5](../model_doc/longt5), [M2M100](../model_doc/m2m_100), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [NLLB](../model_doc/nllb), [NLLB-MOE](../model_doc/nllb-moe), [Pegasus](../model_doc/pegasus), [PEGASUS-X](../model_doc/pegasus_x), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [SeamlessM4T](../model_doc/seamless_m4t), [SeamlessM4Tv2](../model_doc/seamless_m4t_v2), [SwitchTransformers](../model_doc/switch_transformers), [T5](../model_doc/t5), [UMT5](../model_doc/umt5), [XLM-ProphetNet](../model_doc/xlm-prophetnet)
 
 <!--End of the generated tip-->
 

src/transformers/__init__.py

@@ -547,6 +547,10 @@ _import_structure = {
         "SeamlessM4TFeatureExtractor",
         "SeamlessM4TProcessor",
     ],
+    "models.seamless_m4t_v2": [
+        "SEAMLESS_M4T_V2_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "SeamlessM4Tv2Config",
+    ],
     "models.segformer": ["SEGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "SegformerConfig"],
     "models.sew": ["SEW_PRETRAINED_CONFIG_ARCHIVE_MAP", "SEWConfig"],
     "models.sew_d": ["SEW_D_PRETRAINED_CONFIG_ARCHIVE_MAP", "SEWDConfig"],
@@ -2778,6 +2782,17 @@ else:
             "SeamlessM4TTextToUnitModel",
         ]
     )
+    _import_structure["models.seamless_m4t_v2"].extend(
+        [
+            "SEAMLESS_M4T_V2_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "SeamlessM4Tv2ForSpeechToSpeech",
+            "SeamlessM4Tv2ForSpeechToText",
+            "SeamlessM4Tv2ForTextToSpeech",
+            "SeamlessM4Tv2ForTextToText",
+            "SeamlessM4Tv2Model",
+            "SeamlessM4Tv2PreTrainedModel",
+        ]
+    )
     _import_structure["models.segformer"].extend(
         [
             "SEGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -4794,6 +4809,10 @@ if TYPE_CHECKING:
         SeamlessM4TFeatureExtractor,
         SeamlessM4TProcessor,
     )
+    from .models.seamless_m4t_v2 import (
+        SEAMLESS_M4T_V2_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        SeamlessM4Tv2Config,
+    )
     from .models.segformer import SEGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, SegformerConfig
     from .models.sew import SEW_PRETRAINED_CONFIG_ARCHIVE_MAP, SEWConfig
     from .models.sew_d import SEW_D_PRETRAINED_CONFIG_ARCHIVE_MAP, SEWDConfig
@@ -6678,6 +6697,15 @@ if TYPE_CHECKING:
             SeamlessM4TTextToUnitForConditionalGeneration,
             SeamlessM4TTextToUnitModel,
         )
+        from .models.seamless_m4t_v2 import (
+            SEAMLESS_M4T_V2_PRETRAINED_MODEL_ARCHIVE_LIST,
+            SeamlessM4Tv2ForSpeechToSpeech,
+            SeamlessM4Tv2ForSpeechToText,
+            SeamlessM4Tv2ForTextToSpeech,
+            SeamlessM4Tv2ForTextToText,
+            SeamlessM4Tv2Model,
+            SeamlessM4Tv2PreTrainedModel,
+        )
         from .models.segformer import (
             SEGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             SegformerDecodeHead,

src/transformers/models/__init__.py

@@ -185,6 +185,7 @@ from . import (
     rwkv,
     sam,
     seamless_m4t,
+    seamless_m4t_v2,
     segformer,
     sew,
     sew_d,

src/transformers/models/auto/configuration_auto.py

@@ -191,6 +191,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("rwkv", "RwkvConfig"),
         ("sam", "SamConfig"),
         ("seamless_m4t", "SeamlessM4TConfig"),
+        ("seamless_m4t_v2", "SeamlessM4Tv2Config"),
         ("segformer", "SegformerConfig"),
         ("sew", "SEWConfig"),
         ("sew-d", "SEWDConfig"),
@@ -404,6 +405,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("rwkv", "RWKV_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("sam", "SAM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("seamless_m4t", "SEAMLESS_M4T_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("seamless_m4t_v2", "SEAMLESS_M4T_V2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("segformer", "SEGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("sew", "SEW_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("sew-d", "SEW_D_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -642,6 +644,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("rwkv", "RWKV"),
         ("sam", "SAM"),
         ("seamless_m4t", "SeamlessM4T"),
+        ("seamless_m4t_v2", "SeamlessM4Tv2"),
         ("segformer", "SegFormer"),
         ("sew", "SEW"),
         ("sew-d", "SEW-D"),

src/transformers/models/auto/modeling_auto.py

@@ -180,6 +180,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("rwkv", "RwkvModel"),
         ("sam", "SamModel"),
         ("seamless_m4t", "SeamlessM4TModel"),
+        ("seamless_m4t_v2", "SeamlessM4Tv2Model"),
         ("segformer", "SegformerModel"),
         ("sew", "SEWModel"),
         ("sew-d", "SEWDModel"),
@@ -685,6 +686,7 @@ MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("plbart", "PLBartForConditionalGeneration"),
         ("prophetnet", "ProphetNetForConditionalGeneration"),
         ("seamless_m4t", "SeamlessM4TForTextToText"),
+        ("seamless_m4t_v2", "SeamlessM4Tv2ForTextToText"),
         ("switch_transformers", "SwitchTransformersForConditionalGeneration"),
         ("t5", "T5ForConditionalGeneration"),
         ("umt5", "UMT5ForConditionalGeneration"),
@@ -696,6 +698,7 @@ MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES = OrderedDict(
     [
         ("pop2piano", "Pop2PianoForConditionalGeneration"),
         ("seamless_m4t", "SeamlessM4TForSpeechToText"),
+        ("seamless_m4t_v2", "SeamlessM4Tv2ForSpeechToText"),
         ("speech-encoder-decoder", "SpeechEncoderDecoderModel"),
         ("speech_to_text", "Speech2TextForConditionalGeneration"),
         ("speecht5", "SpeechT5ForSpeechToText"),
@@ -1062,6 +1065,7 @@ MODEL_FOR_TEXT_TO_WAVEFORM_MAPPING_NAMES = OrderedDict(
         ("bark", "BarkModel"),
         ("musicgen", "MusicgenForConditionalGeneration"),
         ("seamless_m4t", "SeamlessM4TForTextToSpeech"),
+        ("seamless_m4t_v2", "SeamlessM4Tv2ForTextToSpeech"),
         ("vits", "VitsModel"),
     ]
 )

src/transformers/models/auto/tokenization_auto.py

@@ -346,6 +346,13 @@ else:
                     "SeamlessM4TTokenizerFast" if is_tokenizers_available() else None,
                 ),
             ),
+            (
+                "seamless_m4t_v2",
+                (
+                    "SeamlessM4TTokenizer" if is_sentencepiece_available() else None,
+                    "SeamlessM4TTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
             ("speech_to_text", ("Speech2TextTokenizer" if is_sentencepiece_available() else None, None)),
             ("speech_to_text_2", ("Speech2Text2Tokenizer", None)),
             ("speecht5", ("SpeechT5Tokenizer" if is_sentencepiece_available() else None, None)),

src/transformers/models/seamless_m4t_v2/__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_seamless_m4t_v2": ["SEAMLESS_M4T_V2_PRETRAINED_CONFIG_ARCHIVE_MAP", "SeamlessM4Tv2Config"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_seamless_m4t_v2"] = [
+        "SEAMLESS_M4T_V2_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "SeamlessM4Tv2ForTextToSpeech",
+        "SeamlessM4Tv2ForSpeechToSpeech",
+        "SeamlessM4Tv2ForTextToText",
+        "SeamlessM4Tv2ForSpeechToText",
+        "SeamlessM4Tv2Model",
+        "SeamlessM4Tv2PreTrainedModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_seamless_m4t_v2 import SEAMLESS_M4T_V2_PRETRAINED_CONFIG_ARCHIVE_MAP, SeamlessM4Tv2Config
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_seamless_m4t_v2 import (
+            SEAMLESS_M4T_V2_PRETRAINED_MODEL_ARCHIVE_LIST,
+            SeamlessM4Tv2ForSpeechToSpeech,
+            SeamlessM4Tv2ForSpeechToText,
+            SeamlessM4Tv2ForTextToSpeech,
+            SeamlessM4Tv2ForTextToText,
+            SeamlessM4Tv2Model,
+            SeamlessM4Tv2PreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)