Google's New Transcription Model Cleans Up Your Ums and Your Mind Changes
Gemini 3.5 Transcribe handles 85 languages, strips filler words, and fixes it when you say Tuesday and mean Wednesday. It is already in Gboard on Android and the Gemini app on macOS.
Google has released Gemini 3.5 Transcribe, a speech to text model that does more than write down what you said. It detects and transcribes over 85 languages automatically, removes filler words, formats the text on its own, and quietly fixes self corrections. Say “let’s meet Tuesday, no, Wednesday” and you get Wednesday, not the whole stumble.
The accuracy numbers, measured by the independent outfit Artificial Analysis, are a word error rate of 4.0 percent for live streaming and 2.6 percent for recorded audio. Word error rate is simply the share of words the model gets wrong, so lower is better and 2.6 percent is roughly one slip in forty words. On the multilingual FLEURS benchmark across a set of top languages the figures are higher, 5.50 percent streaming and 5.04 percent recorded, which is the more honest number if you do not speak English. Time to a final transcript is 70 percent faster than Chirp 3, the model it replaces. For recorded audio it can attribute speech to up to three speakers with word level timestamps, and three or more is still experimental.
There are two ways in. The Live API handles real time streaming with sub second delay for things like voice assistants and captions. The Interactions API handles recorded files, meetings, and call logs. Both are in public preview through Google AI Studio, with an enterprise route through the Gemini Enterprise Agent Platform.
What’s actually going on here: conventional speech recognition models transcribe sound to text and stop. This one is a language model doing the transcribing, which is why it can decide that “um” was not part of your sentence and that you changed your mind mid sentence. That is also the risk. A model that edits while it listens is making judgement calls, and for a doctor’s note, a legal deposition, or a research interview you may want the stumbles preserved exactly as spoken. Google does not currently offer a raw verbatim mode as a headline option, so check what you are getting before you build a workflow on it.
What this means for you: if you are on Android, you already have some of this. The Rambler feature in Gboard uses it to turn spoken rambling into formatted text, in selected countries and languages, and you can edit by voice. The Gemini app on macOS has it in English, and Chrome support is coming, which will mean dictating into any web text field. If you build things, this is a plausible replacement for whatever transcription service you pay for now, though it is a preview, so treat it as such. And if you have avoided dictation because cleaning up the output took longer than typing, this is the specific complaint the model was built to answer. Worth one honest test with your own accent before you commit.
Sources
Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5-transcribe/
IBM's Granite 4.2 Is Free, Small Enough to Run Locally, and Built to Think First
Three open models at 3B, 8B, and 30B under an Apache 2.0 licence, plus a 470-million-parameter speech model that transcribes three hours of audio in about a second.
