Real-time speech-to-speech translation

Google Research · 11 months ago · 29

Google developed an end-to-end speech-to-speech translation model that translates spoken audio directly into target language audio while preserving the original speaker's voice. The system achieves a 2-second latency compared to 4-5 seconds in prior cascaded approaches, and currently supports five Latin-based language pairs including English-Spanish, English-German, and English-French. The technology has been deployed in Google Meet on servers and as an on-device feature on Pixel 10 devices to enable real-time cross-language communication.