StreamSpeech

StreamSpeech

Authors: Shaolei Zhang, Qingkai Fang, Shoutao Guo, Zhengrui Ma, Min Zhang, Yang Feng*

Code for ACL 2024 paper "StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning".

💡Highlight:

1. StreamSpeech achieves SOTA performance on both offline and simultaneous speech-to-speech translation.
2. StreamSpeech performs streaming ASR, simultaneous speech-to-text translation and simultaneous speech-to-speech translation via an "All in One" seamless model.
3. StreamSpeech can present intermediate results (i.e., ASR or translation results) during simultaneous translation, offering a more comprehensive low-latency communication experience.

🔥News

• [2025.06.17] We are excited to extend the "All-in-One" feature of StreamSpeech to more general multimodal interactions via developing Stream-Omni. 👉Refer to paper, code & demo, model for more details.

  • Stream-Omni is an GPT-4o-like language-vision-speech chatbot that simultaneously supports interactions across any combination of text, vision, and speech modalities.
  • Stream-Omni can simultaneously produce intermediate textual results (e.g., ASR transcriptions and model responses) during speech interactions, like the advanced voice service of GPT-4o.

• [2024.06.17] Add Web GUI demo, now you can experience StreamSpeech in your local browser.

• [2024.06.05] Paper, code, models and demo of StreamSpeech are available!

⭐Features

Support 8 Tasks

• Offline: Speech Recognition (ASR)✅, Speech-to-Text Translation (S2TT)✅, Speech-to-Speech Translation (S2ST)✅, Speech Synthesis (TTS)✅
• Simultaneous: Streaming ASR✅, Simultaneous S2TT✅, Simultaneous S2ST✅, Real-time TTS✅ under any latency (with one model)

GUI Demo

https://github.com/ictnlp/StreamSpeech/assets/34680227/4d9bdabf-af66-4320-ae7d-0f23e721cd71

Case

Speech Input: example/wavs/common_voice_fr_17301936.mp3

Transcription (ground truth): jai donc lexpérience des années passées jen dirai un mot tout à lheure

Translation (ground truth): i therefore have the experience of the passed years i'll say a few words about that later

StreamSpeech	Simultaneous	Offline
Speech Recognition	jai donc expérience des années passé jen dirairai un mot tout à lheure	jai donc lexpérience des années passé jen dirairai un mot tout à lheure
Speech-to-Text Translation	i therefore have an experience of last years i will tell a word later	so i have the experience in the past years i'll say a word later
Speech-to-Speech Translation
Text-to-Speech Synthesis (incrementally synthesize speech word by word)

⚙Requirements

• Python == 3.10, PyTorch == 2.0.1, Install fairseq & SimulEval

  cd fairseq
  pip install --editable ./ --no-build-isolation
  cd SimulEval
  pip install --editable ./
  

🚀Quick Start

1. Model Download

(1) StreamSpeech Models

Language	UnitY	StreamSpeech (offline)	StreamSpeech (simultaneous)
Fr-En	unity.fr-en.pt [Huggingface] [Baidu]	streamspeech.offline.fr-en.pt [Huggingface] [Baidu]	streamspeech.simultaneous.fr-en.pt [Huggingface] [Baidu]
Es-En	unity.es-en.pt [Huggingface] [Baidu]	streamspeech.offline.es-en.pt [Huggingface] [Baidu]	streamspeech.simultaneous.es-en.pt [Huggingface] [Baidu]
De-En	unity.de-en.pt [Huggingface] [Baidu]	streamspeech.offline.de-en.pt [Huggingface] [Baidu]	streamspeech.simultaneous.de-en.pt [Huggingface] [Baidu]

(2) Unit-based HiFi-GAN Vocoder

Unit config	Unit size	Vocoder language	Dataset	Model
mHuBERT, layer 11	1000	En	LJSpeech	ckpt, config

2. Prepare Data and Config (only for test/inference)

(1) Config Files

Replace /data/zhangshaolei/StreamSpeech in files configs/fr-en/config_gcmvn.yaml and configs/fr-en/config_mtl_asr_st_ctcst.yaml with your local address of StreamSpeech repo.

(2) Test Data

Prepare test data following SimulEval format. example/ provides an example:

• wav_list.txt: Each line records the path of a source speech.
• target.txt: Each line records the reference text, e.g., target translation or source transcription (used to calculate the metrics).

3. Inference with SimulEval

Run these scripts to inference StreamSpeech on streaming ASR, simultaneous S2TT and simultaneous S2ST.

--source-segment-size: set the chunk size (millisecond) to any value to control the latency

--output-asr-translation: whether to output the intermediate ASR and translated text results during simultaneous speech-to-speech translation.

export CUDA_VISIBLE_DEVICES=0

ROOT=/data/zhangshaolei/StreamSpeech # path to StreamSpeech repo
PRETRAIN_ROOT=/data/zhangshaolei/pretrain_models 
VOCODER_CKPT=$PRETRAIN_ROOT/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/g_00500000 # path to downloaded Unit-based HiFi-GAN Vocoder
VOCODER_CFG=$PRETRAIN_ROOT/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/config.json # path to downloaded Unit-based HiFi-GAN Vocoder

LANG=fr
file=streamspeech.simultaneous.${LANG}-en.pt # path to downloaded StreamSpeech model
output_dir=$ROOT/res/streamspeech.simultaneous.${LANG}-en/simul-s2st

chunk_size=320 #ms
PYTHONPATH=$ROOT/fairseq simuleval --data-bin ${ROOT}/configs/${LANG}-en \
    --user-dir ${ROOT}/researches/ctc_unity --agent-dir ${ROOT}/agent \
    --source example/wav_list.txt --target example/target.txt \
    --model-path $file \
    --config-yaml config_gcmvn.yaml --multitask-config-yaml config_mtl_asr_st_ctcst.yaml \
    --agent $ROOT/agent/speech_to_speech.streamspeech.agent.py \
    --vocoder $VOCODER_CKPT --vocoder-cfg $VOCODER_CFG --dur-prediction \
    --output $output_dir/chunk_size=$chunk_size \
    --source-segment-size $chunk_size \
    --quality-metrics ASR_BLEU  --target-speech-lang en --latency-metrics AL AP DAL StartOffset EndOffset LAAL ATD NumChunks DiscontinuitySum DiscontinuityAve DiscontinuityNum RTF \
    --device gpu --computation-aware \
    --output-asr-translation True

You should get the following outputs:

fairseq plugins loaded...
fairseq plugins loaded...
fairseq plugins loaded...
fairseq plugins loaded...
2024-06-06 09:45:46 | INFO     | fairseq.tasks.speech_to_speech | dictionary size: 1,004
import agents...
Removing weight norm...
2024-06-06 09:45:50 | INFO     | agent.tts.vocoder | loaded CodeHiFiGAN checkpoint from /data/zhangshaolei/pretrain_models/unit-based_HiFi-GAN_vocoder/mHuBERT.layer11.km1000.en/g_00500000
2024-06-06 09:45:50 | INFO     | simuleval.utils.agent | System will run on device: gpu.
2024-06-06 09:45:50 | INFO     | simuleval.dataloader | Evaluating from speech to speech.
  0%|                                                                                                                                                                              | 0/2 [00:00<?, ?it/s]
Streaming ASR: 
Streaming ASR: 
Streaming ASR: je
Simultaneous translation: i would
Streaming ASR: je voudrais
Simultaneous translation: i would like to
Streaming ASR: je voudrais soumettre
Simultaneous translation: i would like to sub
Streaming ASR: je voudrais soumettre cette
Simultaneous translation: i would like to submit
Streaming ASR: je voudrais soumettre cette idée
Simultaneous translation: i would like to submit this
Streaming ASR: je voudrais soumettre cette idée à la
Simultaneous translation: i would like to submit this idea to
Streaming ASR: je voudrais soumettre cette idée à la réflexion
Simultaneous translation: i would like to submit this idea to the
Streaming ASR: je voudrais soumettre cette idée à la réflexion de
Simultaneous translation: i would like to submit this idea to the reflection
Streaming ASR: je voudrais soumettre cette idée à la réflexion de lassemblée
Simultaneous translation: i would like to submit this idea to the reflection of
Streaming ASR: je voudrais soumettre cette idée à la réflexion de lassemblée nationale
Simultaneous translation: i would like to submit this idea to the reflection of the
Streaming ASR: je voudrais soumettre cette idée à la réflexion de lassemblée nationale
Simultaneous translation: i would like to submit this idea to the reflection of the national assembly
 50%|███████████████████████████████████████████████████████████████████████████████████                                                                                   | 1/2 [00:04<00:04,  4.08s/it]
Streaming ASR: 
Streaming ASR: 
Streaming ASR: 
Streaming ASR: 
Streaming ASR: jai donc
Simultaneous translation: i therefore
Streaming ASR: jai donc
Streaming ASR: jai donc expérience des
Simultaneous translation: i therefore have an experience
Streaming ASR: jai donc expérience des années
Streaming ASR: jai donc expérience des années passé
Simultaneous translation: i therefore have an experience of last
Streaming ASR: jai donc expérience des années passé jen
Simultaneous translation: i therefore have an experience of last years
Streaming ASR: jai donc expérience des années passé jen dirairai
Simultaneous translation: i therefore have an experience of last years i will
Streaming ASR: jai donc expérience des années passé jen dirairai un mot
Simultaneous translation: i therefore have an experience of last years i will tell a
Streaming ASR: jai donc expérience des années passé jen dirairai un mot tout à lheure
Simultaneous translation: i therefore have an experience of last years i will tell a word
Streaming ASR: jai donc expérience des années passé jen dirairai un mot tout à lheure
Simultaneous translation: i therefore have an experience of last years i will tell a word later
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:06<00:00,  3.02s/it]
2024-06-06 09:45:56 | WARNING  | simuleval.scorer.asr_bleu | Beta feature: Evaluating speech output. Faieseq is required.
2024-06-06 09:46:12 | INFO | fairseq.tasks.audio_finetuning | Using dict_path : /data/zhangshaolei/.cache/ust_asr/en/dict.ltr.txt
Transcribing predictions: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:01<00:00,  1.63it/s]
2024-06-06 09:46:21 | INFO     | simuleval.sentence_level_evaluator | Results:
 ASR_BLEU       AL    AL_CA    AP  AP_CA      DAL  DAL_CA  StartOffset  StartOffset_CA  EndOffset  EndOffset_CA     LAAL  LAAL_CA      ATD   ATD_CA  NumChunks  NumChunks_CA  DiscontinuitySum  DiscontinuitySum_CA  DiscontinuityAve  DiscontinuityAve_CA  DiscontinuityNum  DiscontinuityNum_CA   RTF  RTF_CA
   15.448 1724.895 2913.508 0.425  0.776 1358.812 3137.55       1280.0        2213.906     1366.0        1366.0 1724.895 2913.508 1440.146 3389.374        9.5           9.5             110.0                110.0              55.0                 55.0                 1                    1 1.326   1.326

Logs and evaluation results are stored in $output_dir/chunk_size=$chunk_size:

$output_dir/chunk_size=$chunk_size
├── wavs/
│   ├── 0_pred.wav # generated speech
│   ├── 1_pred.wav 
│   ├── 0_pred.txt # asr transcription for ASR-BLEU tookit
│   ├── 1_pred.txt 
├── config.yaml
├── asr_transcripts.txt # ASR-BLEU transcription results
├── metrics.tsv
├── scores.tsv
├── asr_cmd.bash
└── instances.log # logs of Simul-S2ST

export CUDA_VISIBLE_DEVICES=0

ROOT=/data/zhangshaolei/StreamSpeech # path to StreamSpeech repo

LANG=fr
file=streamspeech.simultaneous.${LANG}-en.pt # path to downloaded StreamSpeech model
output_dir=$ROOT/res/streamspeech.simultaneous.${LANG}-en/simul-s2tt

chunk_size=320 #ms
PYTHONPATH=$ROOT/fairseq simuleval --data-bin ${ROOT}/configs/${LANG}-en \
    --user-dir ${ROOT}/researches/ctc_unity --agent-dir ${ROOT}/agent \
    --source example/wav_list.txt --target example/target.txt \
    --model-path $file \
    --config-yaml config_gcmvn.yaml --multitask-config-yaml config_mtl_asr_st_ctcst.yaml \
    --agent $ROOT/agent/speech_to_text.s2tt.streamspeech.agent.py\
    --output $output_dir/chunk_size=$chunk_size \
    --source-segment-size $chunk_size \
    --quality-metrics BLEU  --latency-metrics AL AP DAL StartOffset EndOffset LAAL ATD NumChunks RTF \
    --device gpu --computation-aware 

export CUDA_VISIBLE_DEVICES=0

ROOT=/data/zhangshaolei/StreamSpeech # path to StreamSpeech repo

LANG=fr
file=streamspeech.simultaneous.${LANG}-en.pt # path to downloaded StreamSpeech model
output_dir=$ROOT/res/streamspeech.simultaneous.${LANG}-en/streaming-asr

chunk_size=320 #ms
PYTHONPATH=$ROOT/fairseq simuleval --data-bin ${ROOT}/configs/${LANG}-en \
    --user-dir ${ROOT}/researches/ctc_unity --agent-dir ${ROOT}/agent \
    --source example/wav_list.txt --target example/source.txt \
    --model-path $file \
    --config-yaml config_gcmvn.yaml --multitask-config-yaml config_mtl_asr_st_ctcst.yaml \
    --agent $ROOT/agent/speech_to_text.asr.streamspeech.agent.py\
    --output $output_dir/chunk_size=$chunk_size \
    --source-segment-size $chunk_size \
    --quality-metrics BLEU  --latency-metrics AL AP DAL StartOffset EndOffset LAAL ATD NumChunks RTF \
    --device gpu --computation-aware 

🎈Develop Your Own StreamSpeech

1. Data Preprocess

• Follow ./preprocess_scripts to process CVSS-C data.

2. Training

[!Note] You can directly use the downloaded StreamSpeech model for evaluation and skip training.

• Follow researches/ctc_unity/train_scripts/train.simul-s2st.sh to train StreamSpeech for simultaneous speech-to-speech translation.
• Follow researches/ctc_unity/train_scripts/train.offline-s2st.sh to train StreamSpeech for offline speech-to-speech translation.
• We also provide some other StreamSpeech variants and baseline implementations.

Model	--user-dir	--arch	Description
Translatotron 2	researches/translatotron	s2spect2_conformer_modified	Translatotron 2
UnitY	researches/translatotron	unity_conformer_modified	UnitY
Uni-UnitY	researches/uni_unity	uni_unity_conformer	Change all encoders in UnitY into unidirectional
Chunk-UnitY	researches/chunk_unity	chunk_unity_conformer	Change the Conformer in UnitY into Chunk-based Conformer
StreamSpeech	researches/ctc_unity	streamspeech	StreamSpeech
StreamSpeech (cascade)	researches/ctc_unity	streamspeech_cascade	Cascaded StreamSpeech of S2TT and TTS. TTS module can be used independently for real-time TTS given incremental text.
HMT	researches/hmt	hmt_transformer_iwslt_de_en	HMT: strong simultaneous text-to-text translation method
DiSeg	researches/diseg	convtransformer_espnet_base_seg	DiSeg: strong simultaneous speech-to-text translation method

[!Tip] The train_scripts/ and test_scripts/ in directory --user-dir give the training and testing scripts for each model. Refer to official repo of UnitY, Translatotron 2, HMT and DiSeg for more details.

3. Evaluation

(1) Offline Evaluation

Follow pred.offline-s2st.sh to evaluate the offline performance of StreamSpeech on ASR, S2TT and S2ST.

(2) Simultaneous Evaluation

A trained StreamSpeech model can be used for streaming ASR, simultaneous speech-to-text translation and simultaneous speech-to-speech translation. We provide agent/ for these three tasks:

• agent/speech_to_speech.streamspeech.agent.py: simultaneous speech-to-speech translation
• agent/speech_to_text.s2tt.streamspeech.agent.py: simultaneous speech-to-text translation
• agent/speech_to_text.asr.streamspeech.agent.py: streaming ASR

Follow simuleval.simul-s2st.sh, simuleval.simul-s2tt.sh, simuleval.streaming-asr.sh to evaluate StreamSpeech.

4. Our Results

Our project page (https://ictnlp.github.io/StreamSpeech-site/) provides some translated speech generated by StreamSpeech, listen to it 🎧.

(1) Offline Speech-to-Speech Translation ( ASR-BLEU: quality )

(2) Simultaneous Speech-to-Speech Translation ( AL: latency | ASR-BLEU: quality )

(3) Simultaneous Speech-to-Text Translation ( AL: latency | BLEU: quality )

(4) Streaming ASR ( AL: latency | WER: quality )

🖋Citation

If you have any questions, please feel free to submit an issue or contact [email protected].

If our work is useful for you, please cite as:

@inproceedings{streamspeech,
      title={StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning}, 
      author={Shaolei Zhang and Qingkai Fang and Shoutao Guo and Zhengrui Ma and Min Zhang and Yang Feng},
      year={2024},
      booktitle = {Proceedings of the 62th Annual Meeting of the Association for Computational Linguistics (Long Papers)},
      publisher = {Association for Computational Linguistics}
}