Google launched Gemini 3.5 Transcribe on Tuesday. The speech-to-text model strips filler words, respects user-supplied vocabularies and tags up to three speakers in a single recording. By turning raw audio into polished, structured text without a human editor, the service cuts the time developers spend cleaning up transcripts for meeting notes, legal filings and more.

Why Google pushed the button now

Google’s audio-processing stack has been evolving for years, but its earlier offering, Chirp 3, still stumbled over pauses, technical terms and speaker switches. Remote work and podcasting have swollen the transcription market, yet many enterprises still rely on manual post-processing or pricey niche vendors. Gemini 3.5 Transcribe answers that friction point: a model that not only recognizes speech but also edits it on the fly.

What the new model actually does

  • Automatic filler-word removal – “um”, “uh” and similar disfluencies disappear as the transcript is generated, leaving a cleaner read-out.
  • Custom vocabularies – Users upload a list of domain-specific words, preventing the model from “correcting” them into generic terms. This matters for fields littered with acronyms, product names or foreign spellings.
  • Speaker attribution – The system identifies up to three distinct voices, assigns each utterance a speaker label and adds a word-level timestamp. Legal teams, medical scribes and journalists can produce time-coded records straight from the source file.
  • Multilingual reach – Google claims improved accuracy across more than 85 languages, a step up from the narrower set of its predecessor.

All of these capabilities sit behind a developer-focused API. Apps request a transcript and receive a JSON payload that already contains the cleaned text, speaker tags and timestamps.

The broader Gemini audio rollout

Gemini 3.5 Transcribe belongs to a wider family of audio models:

  • Gemini 3.5 Live – Optimized for real-time interaction, it handles mid-sentence interruptions better than earlier live models.
  • Gemini 3.5 Live Experimental – A higher-level reasoning variant that narrates its own step-by-step thinking while solving complex tasks.

Both live models are currently available in English on the macOS Gemini app and through the Rambler dictation feature on Android in a handful of regions.

Who stands to gain

Developers can embed a “clean-as-you-speak” pipeline into collaboration tools, content-creation platforms and voice-driven assistants. Enterprises can generate ready-to-publish transcripts, cutting reliance on third-party services that charge per minute and impose strict data-handling clauses. Content creators can publish polished captions without a separate editing pass, speeding video and podcast turnarounds.

Who might feel the pinch

Specialist transcription vendors that charge premium rates for speaker diarization and jargon handling could see demand dip, especially among cost-sensitive startups. The model’s three-speaker limit may also push larger organizations toward dedicated solutions that support more participants in a single call.

Limits and open questions

  • Speaker count – Only three speakers can be distinguished per file; meetings with larger panels will still need external tooling.
  • Language rollout – Multilingual support is announced, but the live models remain English-only, leaving non-English users waiting for full functionality.
  • Privacy – As with any cloud-based speech service, enterprises must weigh Google’s infrastructure convenience against the need to keep sensitive audio off external servers. Google’s terms allow on-premise deployment for certain customers, but that option is not yet public.

What to watch next

Google has hinted at future updates that will raise the speaker ceiling and expand live model coverage to additional languages. Watching the API’s pricing tiers will also be essential; a steep per-minute cost could erode productivity gains for high-volume users. Finally, the adoption curve among developers will reveal whether the convenience of automatic filler removal outweighs any residual errors in niche vocabularies.

Итог: Gemini 3.5 Transcribe превращает утомительную задачу по обработке голосовых записей в один вызов API, предоставляя разработчикам готовый инструмент для получения чистого текста с разметкой спикеров. Его успех зависит от того, насколько быстро Google расширит языковую поддержку, увеличит лимит количества спикеров и решит вопросы конфиденциальности корпоративных данных.