Google's Gemini 3.5 Transcribe adds speaker separation

Previously, feeding an audio file to a general-purpose Gemini model could produce a transcript, but it could not reliably tell you which participant said which line — a serious problem if the point of the document is accountability. Gemini 3.5 Transcribe is built specifically to close that gap, using voice differences to attach speaker labels and returning a start and end time for every word. It also lets you pre-register up to 1,000 technical terms or proper nouns, which should reduce misrecognitions of product names and people's names, and it automatically formats numbers and units, turning spoken "三件" into "3件".

The rollout differed by platform. On the Gemini API it reached general availability on 2026-08-26, while on Vertex AI it arrived as a preview in August 2026 and reached GA in October 2026. Two output modes are offered: VERBATIM, which transcribes speech exactly as spoken and is the only mode that supports speaker separation and word timestamps, and SMART, which removes hesitations and false starts for a cleaner read but cannot be combined with those features.

For companies weighing adoption, the practical question is workflow rather than raw capability. The most common need — knowing who spoke while also cleaning up the text — cannot be met by a single API call today, so the realistic setup is a two-step pipeline: label the speakers first, then polish the text with a separate text model. Whether that added step is worth it depends on how much accuracy matters in a given meeting, especially since speaker separation for three or more people is still labeled experimental.