Gemini 3.5 Transcribe brings diarized, timestamped STT to the API
Google's new speech model does speaker labels, word-level timestamps, and 1,000-term custom vocab across 85+ languages, with a live streaming endpoint too.
Copy markdown
A real STT model on the Gemini API
Gemini 3.5 Transcribe ships in two model IDs — gemini-3.5-transcribe for files and gemini-3.5-transcribe-live for real-time WebSocket streaming. Google reports a 2.6% average word error rate non-streaming and 4.0% streaming, so it's drop-in production transcription without hosting your own Whisper.
Speakers and timestamps, built in
File processing handles speaker diarization for up to 8 voices (3+ experimental) plus word-level timestamps — the two features that turn a raw transcript into usable meeting notes, podcast chapters, or searchable call logs. One tradeoff to know: Google says enabling word-level timestamps slightly degrades accuracy.
Teach it your jargon, in 85+ languages
Custom vocabulary accepts up to 1,000 terms (best results under ~100), so product names, drug names, and code identifiers stop getting mangled. 'Smart transcription' strips filler words and does intent-aware alphanumeric formatting, and the model auto-detects across 85+ languages with live language switching mid-stream.
What to build this weekend
One API now covers live captioning for a call app, a podcast-to-show-notes pipeline with named speakers, or a voice-command front end — the streaming endpoint makes real-time UX viable. It already powers Gboard's Rambler dictation and is headed to Chrome, so it's tuned on messy real-world audio, not benchmark clips.