Google releases Gemini 3.5 Transcribe, a new speech-to-text model for live and non-streaming transcription
Model release ● Confirmed 93% confidence first seen
Google has released Gemini 3.5 Transcribe, an AI speech-to-text model intended for real-time and application-based voice transcription. The coverage describes improvements such as reduced transcription errors and filler-word handling, with availability via Gemini API and AI Studio and rollout across Google products and voice features. Reported performance figures include lower word error rates for both streaming and non-streaming use cases, and the service is positioned as API-based rather than providing open weights.
Decision brief
- What changed
- Google released Gemini 3.5 Transcribe, a new speech-to-text model for both live streaming and non-streaming transcription, and made it available through the Gemini API and Google AI Studio. Coverage says Google is also rolling it into Google voice features and products, with reported improvements in word error rates, speed, jargon handling, and filler-word cleanup versus prior tooling.
- Why it matters
- For leaders responsible for AI product roadmaps, this expands Google’s API-based speech stack into a stronger option for customer support, meeting notes, voice UI, and transcription workflows without requiring self-hosting. The reported gains in accuracy, latency, and multi-language support could improve user experience and operating efficiency, but the API-only delivery model means adoption decisions depend on vendor dependence, pricing, and whether Google’s live versus non-streaming tradeoffs fit the company’s use cases. Assumption: your organization is evaluating or already using speech transcription in products or internal workflows.
- Evidence
- The event is supported across multiple outlets, including Google DeepMind’s launch post, The Verge, Ars Technica, Product Hunt, and MarkTechPost. Core facts are consistent across coverage: Gemini 3.5 Transcribe launched for streaming and non-streaming transcription, is exposed via Google APIs/tools, and is positioned around better accuracy and cleaner transcripts; performance figures mainly trace back to Google’s claims and a cited Artificial Analysis benchmark.
- What remains uncertain
- Open questions include actual pricing, enterprise terms, reliability under specific domain conditions, and whether reported performance holds across accents, noise levels, and industry vocabularies relevant to a given business. It is also not yet clear from the coverage how complete product rollout will be, how the model compares in production against competing vendors on identical workloads, or whether API-only deployment is acceptable for organizations with data residency or self-hosting requirements.
- Monitor next
- Watch for Google to publish pricing, GA timing, and independent head-to-head enterprise benchmark results for Gemini 3.5 Transcribe versus competing speech-to-text APIs.
Analytical support, not advice — assumptions and open questions stated above.