Gemini 3.5 Transcribe separates live dictation from recorded-audio processing

Google introduced Gemini 3.5 Transcribe on 26 August with separate routes for live and prerecorded audio. The streaming version uses the Live API; recorded meetings and call logs use the Interactions API, with speaker attribution and word-level timestamps. The model supports more than 85 automatically detected languages.

Google also describes custom vocabulary and cleanup of filler words or self-corrections. That makes a polished transcript different from a strict record of every spoken word. Applications collecting interviews, evidence or quotations should preserve the original audio and choose their transcription behaviour deliberately. Speaker identification supports up to three people, with larger groups described as experimental. Those limits are more useful for planning a workflow than assuming one transcription mode serves every recording.