Google launches Gemini 3.5 Transcribe with 2.6% error rate

Anyone who has ever tried voice dictation knows the usual headache: one small stutter or a casual "wait, scratch that" usually turns your draft into an embarrassing mess. Traditional speech-to-text models simply write down every stumble without understanding what you actually meant.

Google officially launched Gemini 3.5 Transcribe, a dedicated speech recognition model engineered to solve that exact annoyance by combining pinpoint transcription with context-aware cleanups.

Cutting the filler and fixing mistakes on the fly

According to benchmarks from Artificial Analysis, Gemini 3.5 Transcribe delivers an average Word Error Rate (WER) of 2.6% for non-streaming audio and 4.0% for real-time streaming. Compared to Google's previous Chirp 3 model, it slashes the time to final transcription by 70%, while posting a 5.04% non-streaming WER on the multilingual FLEURS benchmark.

Beyond raw accuracy, the real highlight is its smart transcription mode alongside the classic verbatim output. If you say "let's meet Tuesday, no, wait, Wednesday," the model automatically outputs Wednesday. It scrubs filler sounds like "ums" and "ahs," correctly formats punctuation, and cleanly parses alphanumeric strings like postal codes and order IDs.

Multilingual recognition is another strong suit. It automatically identifies and transcribes more than 85 languages, effortlessly handling speakers who switch between languages like English, Spanish, and Hindi mid-sentence. For recorded audio, it supports speaker attribution with word-level timestamps for up to three people.

From a simple dictation tool to a voice controller

Google is rolling out the model through dual channels. Developers can access gemini-3.5-transcribe-live for sub-second bidirectional streaming via the Live API, or gemini-3.5-transcribe for recorded audio through the Interactions API in Google AI Studio and Google Antigravity.

It is also surfacing directly across user interfaces. On Android, Gboard harnesses the model through the Rambler feature for voice-based drafting and style edits. On macOS, the Gemini app uses it to trigger function calling across other Gemini models, allowing users to analyze screen files or generate images entirely hands-free. A talk-to-type capability across all web fields in Chrome is coming soon.

Ecosystem platforms like LiveKit, LangChain, Vercel, and device maker vivo have already integrated the model into their developer pipelines.

Why it matters

We think Google made a clever architectural shift here by collapsing transcription and initial text cleanup into a single step.

Previously, voice agent builders had to transcribe audio first, then pipe the messy transcript into an LLM just to remove stutters and handle self-corrections. That workflow added avoidable latency and doubled API costs. By baking disfluency cleanup and formatting right into the transcription engine, Gemini 3.5 Transcribe cuts down both system lag and engineering overhead.

For end users, the barrier to voice input drops substantially. You no longer need to speak in rigid, pre-planned sentences; you can simply think out loud, backtrack, and let the model produce a polished note.

That said, limitations remain. Speaker identification beyond three voices is still labeled experimental, meaning crowded boardroom meetings will still pose a challenge. Furthermore, the most powerful capabilities, such as screen-aware voice workflows, are currently tethered to native clients like macOS, so broader utility will depend on how quickly Chrome and Android rollouts mature.