%0 Generic %1 zhang2023google %A Zhang, Yu %A Han, Wei %A Qin, James %A Wang, Yongqiang %A Bapna, Ankur %A Chen, Zhehuai %A Chen, Nanxin %A Li, Bo %A Axelrod, Vera %A Wang, Gary %A Meng, Zhong %A Hu, Ke %A Rosenberg, Andrew %A Prabhavalkar, Rohit %A Park, Daniel S. %A Haghani, Parisa %A Riesa, Jason %A Perng, Ginger %A Soltau, Hagen %A Strohman, Trevor %A Ramabhadran, Bhuvana %A Sainath, Tara %A Moreno, Pedro %A Chiu, Chung-Cheng %A Schalkwyk, Johan %A Beaufays, Françoise %A Wu, Yonghui %D 2023 %K asr conformer %T Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages %U http://arxiv.org/abs/2303.01037 %X We introduce the Universal Speech Model (USM), a single large model that performs automatic speech recognition (ASR) across 100+ languages. This is achieved by pre-training the encoder of the model on a large unlabeled multilingual dataset of 12 million (M) hours spanning over 300 languages, and fine-tuning on a smaller labeled dataset. We use multilingual pre-training with random-projection quantization and speech-text modality matching to achieve state-of-the-art performance on downstream multilingual ASR and speech-to-text translation tasks. We also demonstrate that despite using a labeled training set 1/7-th the size of that used for the Whisper model, our model exhibits comparable or better performance on both in-domain and out-of-domain speech recognition tasks across many languages.

@misc{zhang2023google, abstract = {We introduce the Universal Speech Model (USM), a single large model that performs automatic speech recognition (ASR) across 100+ languages. This is achieved by pre-training the encoder of the model on a large unlabeled multilingual dataset of 12 million (M) hours spanning over 300 languages, and fine-tuning on a smaller labeled dataset. We use multilingual pre-training with random-projection quantization and speech-text modality matching to achieve state-of-the-art performance on downstream multilingual ASR and speech-to-text translation tasks. We also demonstrate that despite using a labeled training set 1/7-th the size of that used for the Whisper model, our model exhibits comparable or better performance on both in-domain and out-of-domain speech recognition tasks across many languages.}, added-at = {2023-11-01T11:43:20.000+0100}, author = {Zhang, Yu and Han, Wei and Qin, James and Wang, Yongqiang and Bapna, Ankur and Chen, Zhehuai and Chen, Nanxin and Li, Bo and Axelrod, Vera and Wang, Gary and Meng, Zhong and Hu, Ke and Rosenberg, Andrew and Prabhavalkar, Rohit and Park, Daniel S. and Haghani, Parisa and Riesa, Jason and Perng, Ginger and Soltau, Hagen and Strohman, Trevor and Ramabhadran, Bhuvana and Sainath, Tara and Moreno, Pedro and Chiu, Chung-Cheng and Schalkwyk, Johan and Beaufays, Françoise and Wu, Yonghui}, biburl = {https://www.bibsonomy.org/bibtex/26e885232b309332c17dd2ad4b0eab118/sivchand}, interhash = {89e2cea98710198e5519ef9e21c29601}, intrahash = {6e885232b309332c17dd2ad4b0eab118}, keywords = {asr conformer}, note = {cite arxiv:2303.01037Comment: 20 pages, 7 figures, 8 tables}, timestamp = {2023-11-01T11:43:20.000+0100}, title = {Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages}, url = {http://arxiv.org/abs/2303.01037}, year = 2023 }