Google introduced Gemini 3.5 Transcribe on August 26, 2026, with separate options for real-time speech and recorded audio. Developers can access it through Google AI Studio and the Gemini Enterprise Agent Platform.
The streaming model, gemini-3.5-transcribe-live, uses the Gemini Live API for two-way speech transcription with sub-second latency. The recorded-audio model, gemini-3.5-transcribe, runs through the Interactions API and returns speaker attribution and word-level timestamps. Google says the model recognizes more than 85 languages and supports custom vocabulary for names and specialized terms.
Google cites Artificial Analysis measurements of 4.0% average word error rate for streaming and 2.6% for recorded audio. Those are benchmark figures for the tested settings, not a guarantee for every accent, noisy recording or language.
The model also handles self-corrections and can remove filler words while formatting dictated text. That is useful for captions and voice-driven writing, though it changes the original phrasing and may not suit work that needs a verbatim transcript. Speaker identification beyond three voices is still experimental, according to Google's documentation.