Google is pushing voice input beyond basic dictation with Gemini 3.5 Transcribe, a new speech-to-text model designed to turn messy, conversational speech into cleaner written text. The technology is already behind Gboard’s Rambler feature and voice controls in the Gemini app for macOS, with Chrome support next on the roadmap.
The important shift here is that Google isn’t treating transcription as a simple exercise in matching sounds to words. Gemini 3.5 Transcribe attempts to interpret how people actually speak, including corrections, filler words and specialised terminology. Say “Tuesday — no, Wednesday,” for example, and the model is designed to understand that Wednesday is the intended result rather than faithfully typing the entire correction.
It can also remove verbal clutter such as “um” and “ah,” automatically format dictated text and accommodate custom vocabulary. That could make it considerably more useful for technical work, names and industry-specific terminology where conventional voice typing often becomes frustrating.
Google says the model automatically detects and transcribes more than 85 languages, including different accents and dialects. Pre-recorded audio can identify and timestamp as many as three speakers, while support beyond three speakers remains experimental.
The company is also making some substantial performance claims. Citing measurements from Artificial Analysis, Google says Gemini 3.5 Transcribe achieves an average word error rate of 4% for streaming transcription and 2.6% for non-streaming workloads. It also claims a 70% improvement in time to final transcription compared with its Chirp 3 model.
On the multilingual FLEURS benchmark, Google reports word error rates of 5.50% in streaming mode and 5.04% for non-streaming transcription. Benchmark numbers are useful indicators, but everyday performance with poor microphones, overlapping conversations and unpredictable background noise will ultimately matter more to users.
Gemini 3.5 Transcribe is also part of Google’s broader attempt to turn voice into an interface for AI rather than merely another keyboard. Function calling allows spoken instructions to be handed to other Gemini models for tasks including image generation and file analysis. The Gemini app’s Speak to Window feature on macOS is one example of that approach.
The model is additionally being used with the microphone input in Google Antigravity, where, with permission, contextual information including screen content and conversation history can help it recognise filenames, documents and other relevant terms.
Chrome could bring the technology to a much wider audience. Google plans to let users dictate directly into web fields, potentially covering everything from messages and posts to longer prompts for Gemini.
Developers can currently access Gemini 3.5 Transcribe in public preview through the Gemini API and Google AI Studio, while enterprise access is available through Google’s agent platform. If its real-world accuracy matches Google’s claims, the bigger story may not be better transcription itself, but voice becoming a genuinely practical way to operate AI-powered software.
