Google is making a serious play for the transcription market with Gemini 3.5 Transcribe, a speech-to-text model that handles over 85 languages without asking users to specify which one they’re speaking. The model strips out filler words, fixes verbal stumbles on the fly, and applies formatting automatically — the kind of cleanup work that used to require a human editor or a second pass.
The numbers are genuinely competitive. Google reports a word error rate of 4.0 percent for streaming audio and 2.6 percent for recorded files, with latency that’s 70 percent lower than its previous-generation Chirp 3 model. For context, that puts it in the same ballpark as specialized transcription services that charge per minute, except this is baked directly into Google’s AI stack.
What’s more interesting is the function calling angle. The model can detect when a user wants something beyond transcription — say, generating an image or running a web search — and hand that task off to another Gemini model. That’s a meaningful shift from treating transcription as an isolated utility to treating it as a front door into the broader AI ecosystem.
The rollout is happening across two interfaces: a Live API for real-time streaming and an Interactions API for recorded audio with speaker attribution and timestamps. It’s already live in Google AI Studio, the Gemini Enterprise Agent Platform, Gboard for Android, and the macOS Gemini app, with Chrome support on the way. The Gboard integration, codenamed “Rambler,” suggests Google sees voice input as a core mobile feature rather than a niche enterprise tool.
Google’s pitch here isn’t just accuracy — it’s invisibility. A transcription layer that cleans up your speech, figures out what you actually meant, and quietly routes requests to the right model could make voice interfaces feel less like a parlor trick and more like a default. Whether developers actually build on it is the real test. Google has shipped impressive speech tech before and watched it languish for lack of ecosystem buy-in — Whisper’s open-source momentum and Deepgram’s developer-first approach prove that raw accuracy alone doesn’t win the market.
