winterspringkle/unispeech

UniSpeech - Large Scale Self-Supervised Learning for Speech

UniSpeech

The family of UniSpeech:

WavLM (arXiv): WavLM: Large-Scale Self-Supervised Pre-training for Full Stack Speech Processing

UniSpeech (ICML 2021): Unified Pre-training for Self-Supervised Learning and Supervised Learning for ASR

UniSpeech-SAT (ICASSP 2022 Submission): Universal Speech Representation Learning with Speaker Aware Pre-Training

ILS-SSL (ICASSP 2022 Submission): Self-Supervised Learning for Speech Recognition with Intermediate Layer Supervision

Model introductions, evaluation results, and model inference instructions are located in their corresponding folders. The source code is here [https://github.com/microsoft/UniSpeech/tree/main/src].

Update

• [HuggingFace Integration] Dec 23, 2021: WavLM models are on HuggingFace .
• [HuggingFace Integration] Octorber 26, 2021: UniSpeech-SAT models are on HuggingFace .
• [Model Release] Octorber 13, 2021: UniSpeech-SAT models are releaseed.
• [HuggingFace Integration] Octorber 11, 2021: UniSpeech models are on HuggingFace .
• [Model Release] June, 2021: UniSpeech v1 models are released.

Pre-trained models

We strongly suggest using our UniSpeech-SAT model for speaker related tasks, since it shows very powerful performance on various speaker related benchmarks.

Model	Pretraining Dataset	Finetuning Dataset	Model
UniSpeech Large EN	Labeled: 1350 hrs en	-	download
UniSpeech Large Multilingual	Labeled: 1350 hrs en + 353 hrs fr + 168 hrs es + 90 hrs it	-	download
Unispeech Large+	Labeled: 1350 hrs en, Unlabeled: 353 hrs fr	-	download
UniSpeech Large+	Labeld: 1350 hrs en, Unlabeled: 168 hrs es	-	download
UniSpeech Large+	Labeled: 1350 hrs en, Unlabeld: 90 hrs it	-	download
UniSpeech Large Multilingual	Labeled: 1350 hrs en + 353 hrs fr + 168 hrs es + 90 hrs it, Unlabeled: 17 hrs ky	-	download
UniSpeech Large+	Labeled: 1350 hrs en, Unlabeled: 353 hrs fr	1 hr fr	download
UniSpeech Large+	Labeld: 1350 hrs en, Unlabeled: 168 hrs es	1 hr es	download
UniSpeech Large+	Labeled: 1350 hrs en, Unlabeld: 90 hrs it	1 hr it	download
UniSpeech Large Multilingual	Labeled: 1350 hrs en + 353 hrs fr + 168 hrs es + 90 hrs it, Unlabeled: 17 hrs ky	1 hr ky	download
UniSpeech-SAT Base	960 hrs LibriSpeech	-	download
UniSpeech-SAT Base+	60k hrs Libri-Light + 10k hrs GigaSpeech + 24k hrs VoxPopuli	-	download
UniSpeech-SAT Large	60k hrs Libri-Light + 10k hrs GigaSpeech + 24k hrs VoxPopuli	-	download
WavLM Base	960 hrs LibriSpeech	-	Azure Storage Google Drive
WavLM Base+	60k hrs Libri-Light + 10k hrs GigaSpeech + 24k hrs VoxPopuli	-	Azure Storage Google Drive
WavLM Large	60k hrs Libri-Light + 10k hrs GigaSpeech + 24k hrs VoxPopuli	-	Azure Storage Google Drive

Universal Representation Evaluation on SUPERB

Downstream Task Performance

We also evaluate our models on typical speaker related benchmarks.

Speaker Verification

Finetune the model with VoxCeleb2 dev data, and evaluate it on the VoxCeleb1

Model	Fix pre-train	Vox1-O	Vox1-E	Vox1-H
ECAPA-TDNN	-	0.87	1.12	2.12
HuBERT large	Yes	0.888	0.912	1.853
Wav2Vec2.0 (XLSR)	Yes	0.915	0.945	1.895
UniSpeech-SAT large	Yes	0.771	0.781	1.669
WavLM large	Yes	0.59	0.65	1.328
WavLM large	No	0.505	0.579	1.176
+Large Margin Finetune and Score Calibration
HuBERT large	No	0.585	0.654	1.342
Wav2Vec2.0 (XLSR)	No	0.564	0.605	1.23
UniSpeech-SAT large	No	0.564	0.561	1.23
WavLM large (New)	No	0.33	0.477	0.984

Large-scale Self-Supervised Speech Representation Learning for Automatic Speaker Verification

Speech Separation

Evaluation on LibriCSS

Model	0S	0L	OV10	OV20	OV30	OV40
Conformer (SOTA)	4.5	4.4	6.2	8.5	11	12.6
UniSpeech-SAT base	4.4	4.4	5.4	7.2	9.2	10.5
UniSpeech-SAT large	4.3	4.2	5.0	6.3	8.2	8.8
WavLM base+	4.5	4.4	5.6	7.5	9.4	10.9
WavLM large	4.2	4.1	4.8	5.8	7.4	8.5

Speaker Diarization

Evaluation on CALLHOME

Model	spk_2	spk_3	spk_4	spk_5	spk_6	spk_all
EEND-vector clustering	7.96	11.93	16.38	21.21	23.1	12.49
EEND-EDA clustering (SOTA)	7.11	11.88	14.37	25.95	21.95	11.84
UniSpeech-SAT large	5.93	10.66	12.9	16.48	23.25	10.92
WavLM Base	6.99	11.12	15.20	16.48	21.61	11.75
WavLm large	6.46	10.69	11.84	12.89	20.70	10.35

License

This project is licensed under the license found in the LICENSE file in the root directory of this source tree. Portions of the source code are based on the FAIRSEQ project.

Microsoft Open Source Code of Conduct

Reference

If you find our work is useful in your research, please cite the following paper:

@inproceedings{Wang2021UniSpeech,
  author    = {Chengyi Wang and Yu Wu and Yao Qian and Kenichi Kumatani and Shujie Liu and Furu Wei and Michael Zeng and Xuedong Huang},
  editor    = {Marina Meila and Tong Zhang},
  title     = {UniSpeech: Unified Speech Representation Learning with Labeled and
               Unlabeled Data},
  booktitle = {Proceedings of the 38th International Conference on Machine Learning,
               {ICML} 2021, 18-24 July 2021, Virtual Event},
  series    = {Proceedings of Machine Learning Research},
  volume    = {139},
  pages     = {10937--10947},
  publisher = {{PMLR}},
  year      = {2021},
  url       = {http://proceedings.mlr.press/v139/wang21y.html},
  timestamp = {Thu, 21 Oct 2021 16:06:12 +0200},
  biburl    = {https://dblp.org/rec/conf/icml/0002WQK0WZ021.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

@article{Chen2021WavLM,
  title   = {WavLM: Large-Scale Self-Supervised  Pre-training   for Full Stack Speech Processing},
  author  = {Sanyuan Chen and Chengyi Wang and Zhengyang Chen and Yu Wu and Shujie Liu and Zhuo Chen and Jinyu Li and Naoyuki Kanda and Takuya Yoshioka and Xiong Xiao and Jian Wu and Long Zhou and Shuo Ren and Yanmin Qian and Yao Qian and Jian Wu and Michael Zeng and Furu Wei},
  eprint={2110.13900},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  year={2021}
}

@article{Chen2021UniSpeechSAT,
  title   = {UniSpeech-SAT: Universal Speech Representation Learning with  Speaker Aware Pre-Training},
  author  = {Sanyuan Chen and Yu Wu and Chengyi Wang and Zhengyang Chen and Zhuo Chen and Shujie Liu and   Jian Wu and Yao Qian and Furu Wei and Jinyu Li and  Xiangzhan Yu},
  eprint={2110.05752},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  year={2021}
}

Contact Information

For help or issues using UniSpeech models, please submit a GitHub issue.

For other communications related to UniSpeech, please contact Yu Wu ([email protected]).