Google AI has released Gemini 3.5 Transcribe, a speech-to-text model available via two distinct API endpoints: the Interactions API for pre-recorded files and the Live API for real-time streaming.

  • The model reports an average word error rate of 2.6% for non-streaming and 4.0% for streaming, as measured by Artificial Analysis.
  • Time to final transcription improves by 70% compared to the previous Chirp 3 model.
  • Automatic detection supports over 85 languages, including mid-sentence code-switching.
  • The Live API offers sub-second latency but lacks speaker diarization and word-level timestamps, which are available in the Interactions API.
  • Smart mode removes disfluencies but cannot be combined with timestamps or diarization.

The split between endpoints allows developers to choose between low-latency streaming for live interfaces or detailed transcription features for post-processing.