Gemini 3.5 Transcribe Removes Filler Words, Covers 85+ Languages
In this article
Google launched Gemini 3.5 Transcribe on August 26, 2026, positioning it as a successor to Chirp 3 with claimed improvements in multilingual performance and word error rates. It is available in public preview through the Gemini API via AI Studio and Antigravity, making it evaluable now rather than on a roadmap.
What Gemini 3.5 Transcribe Does
The model bundles four capabilities that audio pipeline builders typically assemble from separate components. First, automatic disfluency removal strips filler words — specifically "um" and "uh" — without a post-processing step. Second, the model handles automatic text formatting, removing the need for a secondary formatting pass. Third, speaker diarization supports attribution for up to three speakers in pre-recorded audio. Fourth, the model produces word-level timestamps throughout a transcript.
Combined, these reduce the number of inference calls needed to get a clean, structured, attributed transcript from raw audio. Teams building pipeline architectures that extract gains from composition rather than raw model capability will recognise the pattern: fewer hand-offs between specialist models means fewer error-propagation points.
Customised Vocabulary and Language Coverage
The customised vocabulary feature is the most operationally significant capability for domain-specific deployments. Developers can supply a vocabulary list that the model uses to resolve ambiguous transcriptions toward correct spelling — the explicit use case being specialised jargon and unusual proper nouns that a general-purpose transcription model would mishear or autocorrect. This sidesteps the need for a post-processing correction layer keyed to a domain dictionary.
Language support covers more than 85 languages, which Google describes as a particular area of improvement over Chirp 3. No per-language word error rate figures, benchmark datasets, or third-party comparisons were provided at launch. The multilingual advancement claim is qualitative in the source material.
Availability by Platform
| Platform | Feature | Status | Scope |
|---|---|---|---|
| macOS Gemini app | 3.5 Transcribe | Rolling out now | English only |
| Android (Rambler dictation) | 3.5 Transcribe | Rolling out now | Select countries and languages |
| Gemini API (AI Studio / Antigravity) | 3.5 Transcribe | Public preview | Not specified |
| Chrome | 3.5 Transcribe | Coming soon | Not specified |
| Gemini API | 3.5 Live / 3.5 Live Experimental | Not yet launched — date unconfirmed | N/A |
The Live Model Correction
Google initially briefed The Verge on 3.5 Live and 3.5 Live Experimental alongside 3.5 Transcribe. According to that briefing, 3.5 Live improves mid-sentence interruption handling, language recognition, and live visual processing; 3.5 Live Experimental adds real-time step-by-step narration during complex reasoning tasks. After publication, Google corrected the briefing: only 3.5 Transcribe is being announced today, and no new launch date for the Live variants has been provided. No pricing was disclosed for any tier.
What the Numbers Cannot Tell You
Google's claim that 3.5 Transcribe "represents a major advancement" over Chirp 3 is unaccompanied by specific figures. Developers evaluating this against established tools — or against open-source alternatives covered in Chandra OCR 2's open-source benchmark analysis — will need to run domain-specific evaluations rather than rely on launch-day characterisations.
The launch continues a pattern of Google shipping named audio models while Gemini 3.5 Pro, promised for June 2026, remains unreleased. For voice pipeline teams, the API availability is real and testable today. Whether the customised vocabulary feature and speaker diarization hold at scale across the full 85-plus language range is a question developer testing will answer, not the announcement.