Google launches Gemini 3.5 Transcribe for real-time speech-to-text

Google has launched Gemini 3.5 Transcribe, a new speech-to-text model for live voice applications and recorded audio. The company announced the model on August 26, 2026 and made it available in public preview to developers through Google AI Studio and to enterprises through the Gemini Enterprise Agent Platform. It is already used in selected Google products, while a Chrome integration is still forthcoming.

Developers can access two versions. `gemini-3.5-transcribe-live` runs through the Live API and supports continuous, bidirectional streaming with sub-second latency. The separate `gemini-3.5-transcribe` model processes recordings through the Interactions API, adding speaker attribution and word-level timestamps to meetings, calls and other audio. Google says the model automatically detects more than 85 languages, can follow regional accents and custom vocabulary, and can handle corrections made while someone is speaking. It can also remove filler words and format the result rather than producing a strictly literal transcript.

The model is appearing across Google's products in different forms. Rambler on Android turns speech into edited text in Gboard in selected countries and languages. Google Antigravity can use screen context and chat history, with permission, to recognize file names and terms. The Gemini app on macOS combines transcription with commands that can call other Gemini models. Chrome voice typing has been announced but is not available yet.

Google reports average word-error rates of 4.0 percent for streaming and 2.6 percent for non-streaming use, measured by Artificial Analysis. It also says time to a final transcript improved by 70 percent over Chirp 3. Those figures come from the launch material and cited evaluations; they do not guarantee the same accuracy for every language, accent, noisy room or specialist domain.

For makers, the practical gain is one model family for live captions, voice agents and post-call analysis instead of separate speech recognition and cleanup steps. For companies, speaker labels, timestamps and custom terminology can make recorded conversations easier to search and process. The broader consequence is that transcription is shifting from verbatim conversion toward interpretation: the system may polish intent, remove words and use surrounding context. That can save editing time, but it also makes review essential for names, numbers, quotations and regulated records. Teams should test accuracy, consent, retention and access controls before using the preview with sensitive audio.