Google just raised the bar on speech-to-text. According to Google DeepMind, the company has launched Gemini 3.5 Transcribe, which it calls its most precise speech recognition model yet. The pitch is simple: instead of spitting out a raw, error-riddled transcript, it converts audio directly into clean, formatted, ready-to-use text.
What stands out here is the focus on real speech. People stumble, backtrack, and mumble. Most transcription tools choke on that. Google DeepMind reports this model handles background noise, jargon, and disfluency cleanup without a separate editing pass.
What Google actually shipped
Developers can now access the model through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. It comes in two flavors:
- Real-time streaming. Continuous, two-way streaming with sub-second latency for live voice apps, running on
gemini-3.5-transcribe-livevia the Live API. This is the one for voice agents and live captioning. - Pre-recorded processing. Handles meetings, call logs, and recorded audio with speaker labels and word-level timestamps, using
gemini-3.5-transcribethrough the Interactions API.
Consumers have already been using it. Google DeepMind points to voice features like Rambler on Android and the Gemini app on macOS as early proof the model works in the wild.
The capabilities worth knowing
- Smart transcription. It cleans up as it listens. Say “let’s meet Tuesday, no, Wednesday” and it keeps the correction straight. It strips filler words like “ums” and “ahs” and auto-formats the output.
- Function calling. The model can hand off complex jobs, like image generation or file analysis, to other Gemini models through function calls. This is live in the Gemini macOS app for now.
- Lower error rates. As measured by Artificial Analysis, it hits an average Word Error Rate of 4.0% for streaming and 2.6% for non-streaming. It also nails alphanumeric details like postal codes and order IDs, which is where a lot of tools fall apart.
- Custom vocabulary. Feed it your specialized terms or unusual spellings and it adapts, so industry jargon and product names come through right.
- 85-plus languages. It auto-detects and transcribes across regional accents and dialects, no manual language selection needed.
- Multi-speaker ID. For recorded audio, it attributes speech to up to three speakers with timestamps. Support beyond three is still experimental.
Why this matters
Transcription has been a solved-but-annoying problem for years. The output was usually good enough for search but too rough to use directly. You’d get a wall of text full of filler, false starts, and mangled numbers, then spend real time cleaning it. Google DeepMind is betting that the cleanup step is where the value sits, and it’s baking that into the model itself.
That opens practical doors. Post-call analytics pipelines get cleaner input. Voice agents can act on intent instead of parsing garbled commands. Captioning tools get sub-second response. And the function-calling piece hints at where this is headed: voice as a control layer, not just a recording tool. You talk, the model transcribes, decides what you meant, and routes the task to whatever model can do it.
A couple of caveats from Google DeepMind’s own notes. Speaker identification caps cleanly at three people, so busy meetings or panels may still need help. Function calling is limited to the macOS app right now, not the broader API. And the accuracy figures come from Artificial Analysis benchmarks, so your mileage will vary with audio quality and accent mix.
Still, the direction is clear. The competition in speech-to-text has been about shaving error rates by fractions of a percent. Google is trying to change the conversation from “how accurate is the transcript” to “how useful is the output.” For anyone building voice products, that’s the more interesting question.
Developers can start testing both APIs today in Google AI Studio. Full details are available at the original source.