Google's Gemini 3.5 Transcribe turns speech into text in 85+ languages
The new speech recognition model strips filler words, follows self-corrections and formats text on its own. Google reports a word error rate of 4.0% for live streaming and 2.6% for recorded audio.
Google announced Gemini 3.5 Transcribe, a new speech-to-text model. The company presents it as its most precise speech recognition model so far and says it automatically recognises more than 85 languages.
Its distinguishing feature is doing more than a raw transcript. According to Google it strips filler words such as "ums" and "ahs", follows along when a speaker corrects themselves mid-sentence, and formats the text by itself. The aim is to capture what was meant rather than transcribe word for word.
In figures shared by the company, the word error rate is 4.0% for live streaming and 2.6% for recorded audio. Google also says latency is 70% lower than its predecessor, Chirp 3. These are the company's own measurements; no independent comparison is available yet.
The model ships with two interfaces. The Live API handles real-time streaming at very low latency (gemini-3.5-transcribe-live), while the Interactions API processes recorded audio with speaker attribution and timestamps (gemini-3.5-transcribe). Through function calling, the model can hand off tasks such as image generation or web search to other Gemini models.
Gemini 3.5 Transcribe is available through Google AI Studio and the Gemini Enterprise Agent Platform. It also powers Gboard's "Rambler" feature on Android and the Gemini app on macOS, with Chrome support said to be coming soon.
For details see Google's announcement.
