Google's new transcription model cleans up your speech, which is the feature and the risk

Google announced Gemini 3.5 Transcribe on August 26, a speech-to-text model that does not stop at what was said. It drops the ums and ahs, follows a speaker who corrects themselves mid-sentence, and returns formatted text rather than a raw stream of words.

What it does and where it runs

The model comes in two APIs. Real-time streaming runs through the Live API as `gemini-3.5-transcribe-live`, with what Google describes as sub-second latency and bidirectional streaming. Pre-recorded audio goes through the Interactions API as `gemini-3.5-transcribe`, which returns speaker attribution and word-level timestamps. Both are available in the Gemini API through Google AI Studio and the Gemini Enterprise Agent Platform.

Google reports over 85 languages with automatic detection, custom vocabulary for jargon and unusual spellings, and speaker attribution for up to three speakers — beyond three it is marked experimental. The model can also hand work to other Gemini models through function calls, for now only in the Gemini app on macOS.

• Streaming: `gemini-3.5-transcribe-live` through the Live API.
• Recordings: `gemini-3.5-transcribe` through the Interactions API, with timestamps.
• Three speakers attributed reliably; more than three is experimental.
• Already shipping inside Gboard's Rambler, the Gemini app on macOS, Antigravity and AI Studio.

The numbers, and who measured them

Google puts the average word error rate at 4.0 percent for streaming and 2.6 percent for pre-recorded audio, and says time to final transcription improves by 70 percent over its previous model, Chirp 3. On the multilingual FLEURS benchmark it reports 5.50 percent streaming and 5.04 percent non-streaming.

The part worth noticing is the attribution: Google credits Artificial Analysis with the measurement, not its own lab. That does not make the figures independent of the announcement — the vendor still chose which numbers to publish and against which predecessor — but it does mean someone outside Google ran the test, and that is more than most model announcements offer.

Cleanup is not neutral

A transcript that removes filler words and silently resolves "let's meet Tuesday — no, Wednesday" into Wednesday is easier to read and harder to trust. In an interview, a deposition, a medical intake or a newsroom recording, the hesitation and the correction are part of the record, and a model that tidies them away has made an editorial decision on your behalf.

For voice interfaces — dictation, agents, live captions — that is exactly the right default, and it is clearly what this model is built for. For anything that has to stand as evidence of what a person actually said, the raw stream still matters, and Google's post does not say whether one can be kept alongside the cleaned-up text.