NAST-S2X: A Fast and End-to-End Simultaneous Speech-to-Any Translation Model

Features

• 🤖 An end-to-end model without intermediate text decoding
• 💪 Supports offline and streaming decoding of all modalities
• ⚡️ 28× faster inference compared to autoregressive models

Examples

We present an example of French-to-English translation using chunk sizes of 320 ms, 2560 ms, and in offline conditions.

• Generation with chunk sizes of 320 ms and 2560 ms starts generating English translation before the source speech is complete.
• In the examples of simultaneous interpretation, the left audio channel is the input streaming speech, and the right audio channel is the simultaneous translation.

  For a better experience, please wear headphones.

Chunk Size 320ms	Chunk Size 2560ms	Offline

Source Speech Transcript	Reference Text Translation
Avant la fusion des communes, Rouge-Thier faisait partie de la commune de Louveigné.	before the fusion of the towns rouge thier was a part of the town of louveigne

For more examples, please check https://nast-s2x.github.io/.

Performance

• ⚡️ Lightning Fast: 28× faster inference and competitive quality in offline speech-to-speech translation
• 👩‍💼 Simultaneous: Achieves high-quality simultaneous interpretation within a delay of less than 3 seconds
• 🤖 Unified Framework: Support end-to-end text & speech generation in one model

Check Details 👇

Offline-S2S

Simul-S2S	Simul-S2T

Architecture

• Fully Non-autoregressive: Trained with CTC-based non-monotonic latent alignment loss (Shao and Feng, 2022) and glancing mechanism (Qian et al., 2021).
• Minimum Human Design: Seamlessly switch between offline translation and simultaneous interpretation by adjusting the chunk size.
• End-to-End: Generate target speech without target text decoding.

Sources and Usage

Model

We release French-to-English speech-to-speech translation models trained on the CVSS-C dataset to reproduce results in our paper. You can train models in your desired languages by following the instructions provided below.

🤗 Model card

Chunk Size	checkpoint	ASR-BLEU	ASR-BLEU (Silence Removed)	Average Lagging
320ms	checkpoint	19.67	24.90	-393ms
1280ms	checkpoint	20.20	25.71	3330ms
2560ms	checkpoint	24.88	26.14	4976ms
Offline	checkpoint	25.82	-	-

Vocoder
checkpoint

Inference

Before executing all the provided shell scripts, please ensure to replace the variables in the file with the paths specific to your machine.

Offline Inference

• Data preprocessing: Follow the instructions in the document.
• Generate Acoustic Unit: Excute offline_s2u_infer.sh
• Generate Waveform: Excute offline_wav_infer.sh
• Evaluation: Using Fairseq's ASR-BLEU evaluation toolkit

Simultaneous Inference

• We use our customized fork of SimulEval: b43a7c to evaluate the model in simultaneous inference. This repository is built upon the official SimulEval: a1435b and includes additional latency scorers.
• Data preprocessing: Follow the instructions in the document.
• Streaming Generation and Evaluation: Excute streaming_infer.sh

Train your own NAST-S2X

• Data preprocessing: Follow the instructions in the document.
• CTC Pretraining: Excute train_ctc.sh
• NMLA Training: Excute train_nmla.sh

Citing

Please kindly cite us if you find our papers or codes useful.

@inproceedings{
ma2024nonautoregressive,
title={A Non-autoregressive Generation Framework for End-to-End Simultaneous Speech-to-Any Translation},
author={Ma, Zhengrui and Fang, Qingkai and Zhang, Shaolei and Guo, Shoutao and Feng, Yang and Zhang, Min
},
booktitle={Proceedings of ACL 2024},
year={2024},
}

@inproceedings{
fang2024ctcs2ut,
title={CTC-based Non-autoregressive Textless Speech-to-Speech Translation},
author={Fang, Qingkai and Ma, Zhengrui and Zhou, Yan and Zhang, Min and Feng, Yang
},
booktitle={Findings of ACL 2024},
year={2024},
}