Universal Speech Model (sites.research.google)

Article summary

Google has introduced the Universal Speech Model (USM), a state-of-the-art speech recognition model trained on 12 million hours of speech and 28 billion sentences of text, spanning over 300 languages. The model is designed to perform automatic speech recognition on widely-spoken languages as well as under-represented languages. USM has achieved a word error rate (WER) of less than 30% on average across 73 languages and has outperformed the Whisper model in certain scenarios. The model is intended for use in YouTube, such as for closed captions.

Main themes

• speech recognition
• language support
• open-source vs proprietary
• business models
• AI research
• language accessibility
• tech industry trends

What commenters say

• The USM model's improvements over Whisper are marginal and may not be significant enough to justify its use over the open-source alternative.
• The fact that USM is not open-sourced and is behind an API is a major drawback, especially when compared to Whisper's open-source nature.
• The model's ability to recognize under-represented languages is a significant achievement, but the lack of transparency and accessibility may limit its impact.
• The comparison between USM and Whisper is not entirely fair, as Whisper can be fine-tuned for specific languages and may perform better in certain scenarios.
• The business model behind USM, with its limited API access, may be seen as greedy and exploitative, particularly for under-represented communities.
• The development of USM and its potential applications are overshadowed by concerns about Google's business sense and marketing acumen.
• The use of USM for streaming translation and transcription may be limited by its architectural disadvantages compared to other approaches.
• The model's performance on under-represented languages is impressive, but the availability of training data for these languages is a significant challenge.