Google DeepMind · 2026-08-26 · major
Gemini 3.5 Transcribe — Google's speech model cleans up your ums and ahs
Gemini 3.5 Transcribe is Google's new speech-to-text model. It drops filler words, fixes what you correct mid-sentence and formats text as it goes, in 85+ languages. It replaces Chirp 3 and reaches the final transcript 70% faster.
Google's new speech-to-text model writes the sentence you meant, not the one you stumbled through.
Key specs
| Word error rate (streaming) | 4.0% |
|---|---|
| Time to final transcript | 70% faster than Chirp 3 |
Quick facts
| Maker | Google DeepMind |
|---|---|
| Models | gemini-3.5-transcribe, gemini-3.5-transcribe-live |
| Replaces | Chirp 3 |
| Languages | 85+, detected automatically |
| Speakers | Up to 3 labelled (more is experimental) |
| Availability | Public preview in Gemini API and Google AI Studio |
| FLEURS word error rate | 5.50% streaming / 5.04% non-streaming |
What is it?
Gemini 3.5 Transcribe turns messy speech into finished text. It removes filler words like "um" and "ah", follows you when you correct yourself mid-sentence, and applies punctuation and formatting on its own. Two models ship together: gemini-3.5-transcribe for recorded audio and gemini-3.5-transcribe-live for real-time streaming over WebSockets.
How does it work?
Rather than passing audio through a separate speech recogniser and then a clean-up pass, the model reads raw audio and writes polished text in one step. Language detection covers more than 85 languages and runs per utterance, so a bilingual conversation does not need a setting. The recorded-audio model adds word-level timestamps and labels up to three speakers; the live model answers in under a second. Both accept a custom vocabulary for names and jargon, and both can call functions to hand work to other Gemini models.
Why does it matter?
Accuracy on hard audio is where transcription usually falls apart, and Google reports a 4.0% average word error rate streaming and 2.6% non-streaming, plus a 70% cut in time to the final transcript against Chirp 3. Alphanumeric strings — phone numbers, postal codes, account IDs — are called out as a specific target, which is exactly what breaks call-centre and dictation products. Anyone who built on Chirp 3 now has a newer model to move to.
Who is it for?
voice app developers, transcription and call-centre teams
Frequently asked questions
- What is the difference between Gemini 3.5 Transcribe and Gemini 3.5 Transcribe Live?
- Gemini 3.5 Transcribe handles pre-recorded audio and adds word-level timestamps plus speaker labels. Gemini 3.5 Transcribe Live is the streaming version: it sends text back over a WebSocket with sub-second latency, which suits live captions and voice agents. Both cover the same 85+ languages and both support custom vocabulary and function calling.
- What does Gemini 3.5 Transcribe replace?
- Gemini 3.5 Transcribe replaces Chirp 3, the Google transcription model released in 2025. Google reports a 70% improvement in time to the final transcript over Chirp 3, alongside better handling of background noise, jargon and disfluency clean-up. Teams already running on Chirp 3 are the clearest audience for the switch.
- How many speakers can Gemini 3.5 Transcribe tell apart?
- Gemini 3.5 Transcribe labels up to three speakers on pre-recorded audio, with word-level timestamps attached to each. Google marks support beyond three speakers as experimental, so meeting recordings with a large group are not yet a safe fit. Speaker labelling is a feature of the recorded-audio model rather than the live streaming one.
- Where is Gemini 3.5 Transcribe already running?
- Gemini 3.5 Transcribe powers Rambler on Gboard for Android, the Gemini app on macOS, and the microphone in Google Antigravity's prompt box. Google says Chrome is next. Developers reach it in public preview through the Gemini API and Google AI Studio, and enterprises through the Gemini Enterprise Agent Platform.
Try it
Use model id gemini-3.5-transcribe in Google AI Studio or the Gemini API