Google has made Gemini 3.5 Transcribe generally available as a dedicated speech-to-text model, with a separate Transcribe Live endpoint for low-latency streaming.
The non-streaming model can automatically detect more than 85 languages, identify speakers, return word-level timestamps and accept custom vocabulary hints. Google’s Smart transcription mode can also clean up disfluencies and formatting.
There are important limits. File transcription supports up to one hour of audio, falling to 30 minutes when features such as diarization or word-level timestamps are enabled. Live sessions are limited to 10 minutes and do not currently provide diarization or word-level timestamps.
Google is exposing the models through the Gemini API and AI Studio. This is a specialist transcription surface inside the Gemini ecosystem, not simply another capability of the ordinary Gemini chat interface.