Google AI Releases Gemini 3.5 Transcribe: A Speech-to-Text Model Reporting 2.6% Average WER Across 85+ Languages

AI Releases Gemini 3.5 Transcribe: A Speech-to-Text Model Reporting 2.6% Average WER Across 85+ Languages">

Google has released Gemini 3.5 Transcribe , a speech-to-text model for real-time voice interfaces and recorded audio. It ships as two endpoints, not one. gemini-3.5-transcribe handles pre-recorded files through the Interactions API . gemini-3.5-transcribe-live handles bidirectional streaming through the Live API . Google reports average word error rates of 4.0% streaming and 2.6% non-streaming, as measured by Artificial Analysis. Time to final transcription improves 70% over Chirp 3, the previous model. Automatic detection covers more than 85 languages, including mid-sentence code-switching. The split between the two endpoints is the part worth planning around. They do not share the same feature set, limits, or price.

Yes, but API-only. There are no open weights and no self-hosted path. This is a managed-service decision, not an infrastructure one.

The Live API delivers sub-second, continuous transcription. It emits interim_input_transcription for speculative partials while someone is still talking, then input_transcription when the turn finalizes. Audio goes in as raw 16-bit PCM at 16kHz mono, in 100ms chunks. It supports automatic, hybrid, and manual voice-activity detection. Ephemeral tokens let mobile and web clients stream without holding an API key.

The constraints are real. Live sessions cap at 10 minutes of continuous streaming. Speaker diarization is not supported. Word-level timestamps are not supported.

The Interactions API covers what streaming cannot. It offers speaker diarization, word-level start and end offsets, and custom vocabulary biasing. The vocabulary list takes up to 1,000 terms, with best results below 100. Standard requests accept up to one hour of audio. That drops to 30 minutes once diarization or word timestamps are enabled.

Both endpoints expose two modes. verbatim is the default and returns everything, including fillers, repetitions, and false starts. smart removes disfluencies, resolves spoken self-corrections inline, and applies structured formatting.

Google’s own documented example: “Um, so for the meeting, I think we should, uh, invite Alice and, wait no, Bob and Carol.” Verbatim keeps all of it. Smart returns “For the meeting, I think we should invite Bob and Carol.”

Smart mode cannot be combined with word timestamps or diarization. That is the tradeoff to plan around. A readable summary and an auditable transcript are now two different API calls.