Google's new transcription model does not write down what you said — it writes down what you meant, in 85 languages

Google has released Gemini 3.5 Transcribe, a speech-to-text model built on Gemini's audio understanding and now in public preview. It automatically detects more than 85 languages, follows speakers who switch language mid-sentence, separates up to three voices, and emits word-level timestamps. The part that makes it different from ordinary dictation is what Google calls smart transcription: it strips filler words, repairs self-corrections and formats the text as it goes, so the output reads as clean prose rather than a literal record of the audio. It can also be biased towards a custom vocabulary for domain-specific terms, and can call functions to hand off tasks mid-transcription. Google reports a word error rate of 2.6 percent on recorded audio and 4.0 percent streaming, and on the multilingual FLEURS benchmark 5.04 and 5.50 percent respectively, with latency about 70 percent lower than its predecessor Chirp 3. It is available through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, in Gboard's Rambler on Android and the Gemini app on macOS, with Chrome support said to be coming.

✓ Verified · 4 sources

Related stories