Glonce

Google DeepMind ships Gemini 3.5 Transcribe with sub-second streaming

modelsThe Glonce editorsGoogle DeepMind🔥76Glonce rating76Impact75Novelty60Relevance100Surprise45Players95Credibility90
Google DeepMind ships Gemini 3.5 Transcribe with sub-second streaming

Google DeepMind has introduced Gemini 3.5 Transcribe, a new speech-to-text model built for voice interactions, which the company describes as its most precise transcription model yet. Rather than treating recognition and cleanup as separate steps, the model converts raw audio directly into accurate, polished, formatted text, handling self-corrections such as "let's meet Tuesday, no, Wednesday", stripping filler words like "ums" and "ahs", and auto-formatting the output. The model can also delegate complex tasks, such as image generation or file analysis, to other Gemini models via function calls, a capability currently live in the Gemini macOS app.

The model ships across two APIs: a real-time streaming endpoint (gemini-3.5-transcribe-live) built for interactive voice apps through the Live API with sub-second, bidirectional latency, and a pre-recorded audio endpoint (gemini-3.5-transcribe) through the Interactions API that adds speaker attribution and word-level timestamps for meetings, call logs, and similar recordings. Speaker attribution with timestamps is accurate for up to three speakers in pre-recorded audio, with support beyond three still experimental.

On accuracy, Google DeepMind cites figures from Artificial Analysis: an average word error rate of 4.0% for streaming use cases and 2.6% for non-streaming ones, plus a 70% improvement in time to final transcription compared to the company's previous transcription model, Chirp 3. On the FLEURS multilingual benchmark, the model scores 5.50% WER in streaming mode and 5.04% WER non-streaming, which Google DeepMind presents as an improvement over Chirp 3. The model automatically detects and transcribes more than 85 languages, including regional accents and dialects, and can be adapted to custom vocabulary such as specialized jargon and unusual spellings.

Consumers are already using the underlying model through the Rambler feature on Android and the Gemini app on macOS, where it also powers voice commands paired with on-screen context. In Google Antigravity, the model combines screen context and chat history, with user permission, to improve transcription accuracy across file names, agent output and open documents. In Google AI Studio's Build mode, it lets users vibe-code apps by voice. Support for talking to type in any web field is coming soon to Chrome.

Developers can access Gemini 3.5 Transcribe now, in public preview, through the Gemini API in Google AI Studio and through Google Antigravity; enterprises can reach it through the Gemini Enterprise Agent Platform, with support for Gemini Enterprise for Customer Experience coming soon. A set of developer platforms built on the Gemini Live API, including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents, already let developers build voice-driven interfaces on top of it without managing the underlying real-time media streaming infrastructure themselves. Companies including Vivo, Intellitek Health and Lingopal gave positive early feedback, citing latency, accuracy and language coverage.

Key facts

  • Gemini 3.5 Transcribe achieves a 4.0% word error rate for streaming and 2.6% for non-streaming transcription, as measured by Artificial Analysis.
  • Time to final transcription improves by 70% over Google's previous model, Chirp 3, according to the same measurement.
  • On the FLEURS benchmark the model scores 5.50% WER streaming and 5.04% WER non-streaming, an improvement over Chirp 3.
  • The model auto-detects over 85 languages and attributes speech to up to three speakers with timestamps in pre-recorded audio.
  • It is available now via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, and already powers Rambler on Android and the Gemini app on macOS.

Why it matters

Speech-to-text has long traded off accuracy against latency and formatting quality, forcing developers to bolt on separate cleanup steps after transcription. Gemini 3.5 Transcribe folds disfluency removal and formatting into the transcription step itself, and pairs that with a real streaming API fast enough for live voice agents. The 70% cut in time to final transcription, alongside lower error rates than Google's own prior model, is the kind of latency and accuracy jump that determines whether voice interfaces feel usable in real time rather than laggy.

Who it affects

Developers building voice agents, live captioning tools, or post-call analytics pipelines gain a model accessible through the Gemini API in Google AI Studio, with an enterprise route through the Gemini Enterprise Agent Platform. Consumers already encounter the model indirectly through Rambler on Android and the Gemini app on macOS, and soon through voice dictation in Chrome. Platforms that manage real-time voice infrastructure, including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents, can build on top of it via the Gemini Live API.

How to use it

Real-time, interactive use cases go through the Live API with the gemini-3.5-transcribe-live model for sub-second, bidirectional streaming. Pre-recorded audio, meetings and call logs go through the Interactions API with the gemini-3.5-transcribe model, which adds speaker attribution and word-level timestamps. Both are in public preview for developers now; enterprise access comes via the Gemini Enterprise Agent Platform, with Gemini Enterprise for Customer Experience support coming soon. No pricing details are given in the announcement.

How solid is it

The accuracy and latency figures, the 4.0%/2.6% WER split, the 70% latency improvement, and the FLEURS scores, are all attributed to third-party measurement firm Artificial Analysis rather than left as unverified internal claims, and are benchmarked directly against Google's own previous model, Chirp 3. That gives the comparison a fixed baseline, though the announcement does not include any comparison against competitors' transcription models.

Risks and caveats

All performance claims come from Google DeepMind's own blog post citing Artificial Analysis measurements; independent replication is not part of the source material. Multi-speaker attribution is explicitly flagged as experimental beyond three speakers. The Chrome integration and Gemini Enterprise for Customer Experience support are both described only as "coming soon" with no committed date. Antigravity's context-aware transcription draws on screen content and chat history, gated behind user permission, which is worth noting for anyone evaluating what the model has access to.