saved
Intelligent transcription with Gemini 3.5 Transcribe
gist
Google announced Gemini 3.5 Transcribe as its most precise speech-to-text model, already in the Gemini app and Android Rambler, and now in public preview on the Gemini API. Developers get two endpoints: gemini-3.5-transcribe-live for sub-second bidirectional streaming, and gemini-3.5-transcribe for recorded audio with speaker labels and word timestamps. Versus Chirp 3, Google cites Artificial Analysis word-error rates of 4.0% streaming and 2.6% offline, plus 70% faster time to final transcription, more than 85 languages, custom vocabulary, and disfluency cleanup. Diarization is limited to three speakers.
ideas
- Two APIs, two jobs. gemini-3.5-transcribe-live streams with sub-second latency; gemini-3.5-transcribe labels speakers and timestamps words on recorded audio.
- Smart transcription is the product claim. The model is sold on self-corrections, filler-word removal, auto-formatting, and custom vocabulary, not raw recognition alone.
- The scoreboard is versus Chirp 3. Google cites Artificial Analysis WER of 4.0% streaming and 2.6% offline, 70% faster time to final transcription, and better FLEURS multilingual scores.
- Diarization is capped. Pre-recorded attribution covers up to three speakers; more than three is experimental.
- It is already in Google surfaces. Rambler on Gboard, Antigravity with screen context, and the Gemini macOS app ship now; Chrome talk-to-type is coming. Developer and enterprise access is public preview.
quotes
“Today, we’re introducing Gemini 3.5 Transcribe, our most precise speech-to-text model yet”
“Unlike conventional speech recognition models that struggle with background noise, complex jargon, and disfluency cleanup”
“As measured by Artificial Analysis, achieves an average Word Error Rate (WER) of 4.0% for streaming and 2.6% for non-streaming use-cases.”
“As measured by Artificial Analysis, time to final transcription, for example, improves by 70%.”