Intelligent transcription with Gemini 3.5 Transcribe

What it claims

Gemini 3.5 Transcribe is Google's new speech-to-text model and the stated successor to Chirp 3. Google says it beats Chirp 3 on word error rate and on latency: time to final transcription improves by 70%, as measured by Artificial Analysis. It ships as two model IDs on two different APIs:

• gemini-3.5-transcribe-live for streaming.
• gemini-3.5-transcribe for batch.

So streaming and batch transcription need different integration code, not just a different parameter.

Evidence

All figures come from Google's announcement. The page gives no per-dataset breakdown, no noise conditions and no list of languages.

• Latency: time to final transcription is 70% faster than Chirp 3, per Artificial Analysis.
• Accuracy claims without numbers: the page says the model handles noisy environments and alphanumeric entities such as postal codes and order IDs, but gives no figures for either.

How to use it

• Streaming uses the Live API with gemini-3.5-transcribe-live. It is bidirectional, with sub-second latency.
• Pre-recorded audio uses the Interactions API with gemini-3.5-transcribe. This is the only path that returns speaker attribution and word-level timestamps.
• Developer access is in public preview in the Gemini API, through Google AI Studio and Google Antigravity.
• Enterprise access is in public preview through Gemini Enterprise Agent Platform. It is "coming soon" to Gemini Enterprise for Customer Experience.
• Framework integrations: if you already use Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel or Vision Agents, the page says these support the Live API.
• Not on the page: pricing, request or response schemas, and code samples.

Limits

• Preview status: both developer and enterprise access are public preview only. Plan for changes before GA.
• Speaker identification: reliable up to three speakers. More than three is labelled experimental, and diarization is described only for pre-recorded audio.
• Function calling: delegating to other Gemini models, for tasks like image generation and file analysis, is "currently available in the Gemini macOS app". The page does not say it is available to API developers.
• Context-aware accuracy: in Antigravity, accuracy draws on screen context and chat history, with user permission. Don't expect that accuracy on file names and similar terms from the plain API unless you supply custom vocabulary.
• Consumer availability:
• Gemini macOS app: English only.
• Rambler on Gboard: select countries and languages only.
• Chrome: not yet shipped.
• Unknown cost of cleanup: filler removal, self-correction handling and auto-formatting are built in. The page does not say whether they can be turned off for verbatim transcripts, which matters for legal or clinical use.