Google introduced Gemini 3.5 Transcribe on August 26, 2026, in public preview for developers and enterprise users. The release covers both live speech and recorded audio. The most important implementation choice is whether the application needs responsive partial text, a stable transcript, or edited dictation that prioritizes readability.
Choose the appropriate audio path
The launch names gemini-3.5-transcribe-live for streaming and gemini-3.5-transcribe for prerecorded processing. It describes the latter with speaker attribution and word-level timestamps. Preview availability should not be relabeled general availability.
A real-time caption interface and an offline call-analysis pipeline have different acceptance criteria. Captions must remain responsive while someone speaks. Call analysis can wait for a stable record but needs reliable attribution and evidence for downstream summaries. Evaluate the service on the criterion the application actually depends on.
Partial text is not a final business fact
The live-transcription guide distinguishes interim hypotheses from finalized text. Use partial results to make the interface feel responsive, then reconcile the display when the final segment arrives. Do not treat a provisional recognition as an authorized instruction to mutate another system.
This matters for addresses, order identifiers, and self-corrections. A speaker can revise a number before finishing the sentence. The application needs to know when it has a usable record and when it should ask for confirmation. Transcript completion and authorization are separate decisions.
Readable dictation and faithful transcription are different goals
Google describes smart cleanup and language support in the announcement, while the documentation exposes verbatim and smart modes. The recorded-audio guide is the relevant implementation reference for that separate path.
A cleaned draft can be useful for a message or note. An evidentiary record may need to preserve hesitations, corrections, and the original audio. Choose the mode deliberately and retain the source when the workflow depends on what was actually said. Avoid presenting edited prose as an exact quotation.
Evaluate the failures that change outcomes
Build an audio set with the expected languages, accents, background noise, overlapping speakers, and domain terms. Review entity accuracy as well as overall recognition. A low aggregate error rate can hide a mistake in the one number that determines an action.
Test disconnections and the end of the audio stream so segments are not silently omitted or committed twice. Establish consent, access, retention, and deletion rules for recordings and transcripts. Gemini 3.5 Transcribe is a useful new candidate; a reliable voice workflow still needs those decisions around the model.