Final November, we introduced the 1,000 Languages Initiative, an bold dedication to construct a machine studying (ML) mannequin that may assist the world’s one thousand most-spoken languages, bringing larger inclusion to billions of individuals across the globe. Nonetheless, a few of these languages are spoken by fewer than twenty million individuals, so a core problem is easy methods to assist languages for which there are comparatively few audio system or restricted out there knowledge.
As we speak, we’re excited to share extra concerning the Common Speech Mannequin (USM), a essential first step in the direction of supporting 1,000 languages. USM is a household of state-of-the-art speech fashions with 2B parameters educated on 12 million hours of speech and 28 billion sentences of textual content, spanning 300+ languages. USM, which is to be used in YouTube (e.g., for closed captions), can carry out computerized speech recognition (ASR) not solely on widely-spoken languages like English and Mandarin, but additionally on under-resourced languages like Amharic, Cebuano, Assamese, and Azerbaijani to call just a few. In “Google USM: Scaling Automated Speech Recognition Past 100 Languages”, we reveal that using a big unlabeled multilingual dataset to pre-train the encoder of the mannequin and fine-tuning on a smaller set of labeled knowledge allows us to acknowledge under-represented languages. Furthermore, our mannequin coaching course of is efficient at adapting to new languages and knowledge.
 |
| A pattern of the languages that USM helps. |
Challenges in present ASR
To perform this bold aim, we have to tackle two vital challenges in ASR.
First, there’s a lack of scalability with standard supervised studying approaches. A elementary problem of scaling speech applied sciences to many languages is acquiring sufficient knowledge to coach high-quality fashions. With standard approaches, audio knowledge must be both manually labeled, which is time-consuming and dear, or collected from sources with pre-existing transcriptions, that are tougher to seek out for languages that lack extensive illustration. In distinction, self-supervised studying can leverage audio-only knowledge, which is on the market in a lot bigger portions throughout languages. This makes self-supervision a greater strategy to perform our aim of scaling throughout tons of of languages.
One other problem is that fashions should enhance in a computationally environment friendly method whereas we develop the language protection and high quality. This requires the training algorithm to be versatile, environment friendly, and generalizable. Extra particularly, such an algorithm ought to be capable of use giant quantities of information from a wide range of sources, allow mannequin updates with out requiring full retraining, and generalize to new languages and use circumstances.
Our strategy: Self-supervised studying with fine-tuning
USM makes use of the usual encoder-decoder structure, the place the decoder could be CTC, RNN-T, or LAS. For the encoder, USM makes use of the Conformer, or convolution-augmented transformer. The important thing element of the Conformer is the Conformer block, which consists of consideration, feed-forward, and convolutional modules. It takes as enter the log-mel spectrogram of the speech sign and performs a convolutional sub-sampling, after which a collection of Conformer blocks and a projection layer are utilized to acquire the ultimate embeddings.
Our coaching pipeline begins with step one of self-supervised studying on speech audio overlaying tons of of languages. Within the second elective step, the mannequin’s high quality and language protection could be improved by means of an extra pre-training step with textual content knowledge. The choice to include the second step is determined by whether or not textual content knowledge is on the market. USM performs greatest with this second elective step. The final step of the coaching pipeline is to fine-tune on downstream duties (e.g., ASR or computerized speech translation) with a small quantity of supervised knowledge.
For step one, we use BEST-RQ, which has already demonstrated state-of-the-art outcomes on multilingual duties and has confirmed to be environment friendly when utilizing very giant quantities of unsupervised audio knowledge.
Within the second (elective) step, we used multi-objective supervised pre-training to include data from further textual content knowledge. The mannequin introduces an extra encoder module to take textual content as enter and extra layers to mix the output of the speech encoder and the textual content encoder, and trains the mannequin collectively on unlabeled speech, labeled speech, and textual content knowledge.
Within the final stage, USM is fine-tuned on the downstream duties. The general coaching pipeline is illustrated under. With the data acquired throughout pre-training, USM fashions obtain good high quality with solely a small quantity of supervised knowledge from the downstream duties.
 |
| USM’s general coaching pipeline. |
Key outcomes
Efficiency throughout a number of languages on YouTube Captions
Our encoder incorporates 300+ languages by means of pre-training. We reveal the effectiveness of the pre-trained encoder by means of fine-tuning on YouTube Caption’s multilingual speech knowledge. The supervised YouTube knowledge contains 73 languages and has on common lower than three thousand hours of information per language. Regardless of restricted supervised knowledge, the mannequin achieves lower than 30% phrase error price (WER; decrease is best) on common throughout the 73 languages, a milestone we now have by no means achieved earlier than. For en-US, USM has a 6% relative decrease WER in comparison with the present inner state-of-the-art mannequin. Lastly, we examine with the not too long ago launched giant mannequin, Whisper (large-v2), which was educated with greater than 400k hours of labeled knowledge. For the comparability, we solely use the 18 languages that Whisper can efficiently decode with decrease than 40% WER. Our mannequin has, on common, a 32.7% relative decrease WER in comparison with Whisper for these 18 languages.
 |
| USM helps all 73 languages within the YouTube Captions’ Check Set and outperforms Whisper on the languages it might assist with decrease than 40% WER. Decrease WER is best. |
Generalization to downstream ASR duties
On publicly out there datasets, our mannequin exhibits decrease WER in comparison with Whisper on CORAAL (African American Vernacular English), SpeechStew (en-US), and FLEURS (102 languages). Our mannequin achieves decrease WER with and with out coaching on in-domain knowledge. The comparability on FLEURS reviews the subset of languages (62) that overlaps with the languages supported by the Whisper mannequin. For FLEURS, USM with out in-domain knowledge has a 65.8% relative decrease WER in comparison with Whisper and has a 67.8% relative decrease WER with in-domain knowledge.
 |
| Comparability of USM (with or with out in-domain knowledge) and Whisper outcomes on ASR benchmarks. Decrease WER is best. |
Efficiency on computerized speech translation (AST)
For speech translation, we fine-tune USM on the CoVoST dataset. Our mannequin, which incorporates textual content through the second stage of our pipeline, achieves state-of-the-art high quality with restricted supervised knowledge. To evaluate the breadth of the mannequin’s efficiency, we phase the languages from the CoVoST dataset into excessive, medium, and low based mostly on useful resource availability and calculate the BLEU rating (larger is best) for every phase. As proven under, USM outperforms Whisper for all segments.
 |
| CoVoST BLEU rating. Larger BLEU is best. |
Towards 1,000 languages
The event of USM is a essential effort in the direction of realizing Google’s mission to prepare the world’s info and make it universally accessible. We imagine USM’s base mannequin structure and coaching pipeline comprise a basis on which we are able to construct to develop speech modeling to the subsequent 1,000 languages.
Study Extra
Take a look at our paper right here. Researchers can request entry to the USM API right here.
Acknowledgements
We thank all of the co-authors for contributing to the undertaking and paper, together with Andrew Rosenberg, Ankur Bapna, Bhuvana Ramabhadran, Bo Li, Chung-Cheng Chiu, Daniel Park, Françoise Beaufays, Hagen Soltau, Gary Wang, Ginger Perng, James Qin, Jason Riesa, Johan Schalkwyk, Ke Hu, Nanxin Chen, Parisa Haghani, Pedro Moreno Mengibar, Rohit Prabhavalkar, Tara Sainath, Trevor Strohman, Vera Axelrod, Wei Han, Yonghui Wu, Yongqiang Wang, Yu Zhang, Zhehuai Chen, and Zhong Meng.
We additionally thank Alexis Conneau, Min Ma, Shikhar Bharadwaj, Sid Dalmia, Jiahui Yu, Jian Cheng, Paul Rubenstein, Ye Jia, Justin Snyder, Vincent Tsang, Yuanzhong Xu, Tao Wang for helpful discussions.
We recognize helpful suggestions and assist from Eli Collins, Jeff Dean, Sissie Hsiao, Zoubin Ghahramani. Particular due to Austin Tarango, Lara Tumeh, Amna Latif, and Jason Porta for his or her steering round Accountable AI practices. We thank Elizabeth Adkison, James Cokerille for assist with naming the mannequin, Tom Small for the animated graphic, Abhishek Bapna for editorial assist, and Erica Moreira for useful resource administration . We thank Anusha Ramesh for suggestions, steering, and help with the publication technique, and Calum Barnes and Salem Haykal for his or her helpful partnership.
