Google released Gemini 3.5 Transcribe, a speech model built for voice agents rather than captions -- and its defining feature is that it does not transcribe what you actually said. The model deliberately removes filler words, resolves speakers' mid-sentence self-corrections, auto-formats the output, and can hand work off to other Gemini models through function calls. For an agent that needs clean intent, this is an upgrade. For anyone who needs a record, it is a problem.
Key facts
- Two endpoints:
gemini-3.5-transcribe-livefor real-time bidirectional streaming, andgemini-3.5-transcribefor pre-recorded audio with speaker attribution and word-level timestamps. - Time to a final transcript improves roughly 70% over Chirp 3, its predecessor; 85+ languages with automatic detection and mid-stream language switching; diarization capped at 3 speakers.
- Announced August 26, 2026 by Diego Melendo Casado, Senior Director of Engineering for Gemini Audio, and Luke Leonhard, Chief of Staff for Gemini Audio.
- Primary source: Google, "Gemini 3.5 Transcribe"; Live transcription docs.
The hook. Every speech-recognition launch for a decade has been a word-error-rate contest. This one quietly changes what the output is supposed to be.
Background. Most production voice assistants are not single models. They are cascades: speech-to-text turns audio into words, a language model decides what to do, text-to-speech says the answer. Each stage adds delay, and the delay is what makes an assistant feel slow or fast. Our explainers on automatic speech recognition and full-duplex speech models cover the two competing designs. We have also covered Cohere's open Arabic speech model on the open-weights side of this market.
What they did. Google shipped one model family behind two endpoints. The live endpoint streams bidirectionally through the Live API with sub-second latency for interactive voice apps; the batch endpoint handles recorded audio with speaker labels and word-level timing. Both are in Google AI Studio, and the live one is also in the Gemini Enterprise Agent Platform.
The agent framing is literal rather than promotional. Google's post says the model "can delegate complex tasks (such as image generation and file analysis) to other Gemini models via function calls," a capability currently surfaced in the Gemini macOS app. That makes transcription a routing component sitting inside an agent's tool-use loop rather than a terminal step that returns a string.
How it works. Think of the difference between a court stenographer and a good executive assistant. The stenographer writes down every "um," every false start, every reversal, because the record is the point. The assistant hands you a note that says the meeting is Wednesday, because you asked for Tuesday and then corrected yourself, and the useful output is your intent, not your transcript. Gemini 3.5 Transcribe is the assistant. Downstream, that means the language model receives shorter, cleaner, already-resolved text -- fewer tokens to process and fewer chances to misparse a correction.
Why it matters. On accuracy, Google cites third-party measurements from Artificial Analysis rather than self-reporting: on the multilingual FLEURS suite the model misses roughly one word in twenty when streaming and slightly fewer when processing a recording. Its headline figures on an unspecified mix are better still, which is exactly why any cross-vendor comparison should use the FLEURS pair rather than the headline. The number that governs how a voice agent feels, though, is the roughly 70% cut in time-to-final-transcript against Chirp 3. Latency, not accuracy, is what makes an assistant seem present.
There is also a distribution detail worth noticing: the model was already live in consumer products -- Rambler in Gboard on Android, the Gemini app on macOS -- before developers got access, and Google names Antigravity and Chrome as surfaces getting context-aware dictation. This is productization of something already battle-tested, which de-risks the latency claim but means the news is availability rather than capability.
The honest caveat. Two, and they are both structural. First, smart transcription editorializes the record. Post-call analytics, legal discovery, medical documentation and compliance recording all treat disfluency as data -- how someone hesitated is often the finding. Google's post offers no discussion of a verbatim mode. Second, three-speaker diarization is thin against dedicated pipelines for exactly the multi-party call scenario the launch advertises, and anything past three speakers is explicitly experimental. Pricing appears nowhere in the launch post.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)