Universal Speech Model

The Universal Speech Model (USM) is Google's large-scale speech recognition and translation model announced in March 2023. Trained on 12 million hours of speech and 28 billion sentences of text spanning more than 300 languages, it was designed to bring automatic speech recognition to the world's most under-resourced languages.

USM powers downstream systems including YouTube captions and multilingual speech translation, and the research demonstrated strong performance on low-resource languages with minimal supervised data. The model family scales from 2 billion parameters upward.

Google published the research but did not release public weights. USM's training recipe and architecture remain influential in large-scale multilingual ASR research.