Gemini 3.5 Transcribe writes down what you meant, not what you said

Google spent Wednesday rolling out a transcription model, which sounds like the least interesting launch in a month full of agents and mystery models. It might be the biggest one.

@Google introduced Gemini 3.5 Transcribe as "our most precise speech-to-text model yet": it transcribes more than 85 languages, strips out the ums and ahs, and quietly fixes the spots where you correct yourself mid-sentence. @GoogleAIStudio pitched the developer version as "designed for intelligent voice interactions." Within the hour, @testingcatalog confirmed the rollout, noting "auto-detection of 85+ languages out of the box" and that the model was already live in Google AI Studio, the API, and the Gemini desktop app for macOS.

The benchmark numbers are strong. Artificial Analysis measured a 4.0 percent word error rate for live streaming and 2.6 percent for recorded audio, and Google says latency dropped 70 percent compared with Chirp 3, the model this one replaces. But raw accuracy has been close to solved for a couple of years. The real change is in the job description. Every transcription engine until now has been a stenographer. This one is an editor.

The model decides what you meant

Tell it "move the meeting to Tuesday, no, wait, Wednesday" and the transcript says Wednesday. Say um a dozen times in a rambling voicemail and none of them survive. Google calls this smart transcription: the model listens for what you were trying to say and hands back text that reads like you typed it carefully. The rest of the spec sheet:

• Automatic detection of more than 85 languages, including speakers who switch languages mid-sentence
• Speaker labels and word-level timestamps for up to three speakers in recorded audio, with more than three still experimental
• Custom vocabulary biasing, so it can learn your product names, industry jargon, and the odd spelling of a customer's name

My read: this is the right product, and it should also make you a little careful. For dictation it is exactly what people have wanted since the first time they watched a phone type "um so basically." Records are a different matter. A transcript with the hesitations removed and the self-corrections resolved is an interpretation of the call, and when the model resolves one wrong, the result still reads clean and confident, with nothing to mark where the judgment happened. Verbatim or cleaned is now a real decision. If you log calls for disputes, compliance, or training, make it on purpose.

The Chrome line is the one to watch

The model already powers Rambler, the voice feature in Gboard on Android, and dictation in the Gemini app on macOS. Google also says talk-to-type is coming to Chrome, so you will be able to speak into any text field on any website.

That is the part I would not scroll past. Dictation apps are a niche. Every form field in the world's most used browser is not. Once talking becomes a normal way to fill in a support ticket or an email draft, far more business information starts life as speech, and the systems that can catch it cleanly get more valuable.

What this is worth to a working business

Phone calls are the last big pile of information in most small businesses that never makes it into a system. A customer explains the problem to whoever answers, then explains it again to the tech who shows up. A quote gets talked through on a call and rebuilt later from memory. Transcription with speaker labels, timestamps, and a vocabulary that knows your part numbers changes what a recorded call is. It stops being an audio file nobody replays and becomes a searchable record you can build on. An intake call can draft its own job ticket. A month of calls can be searched like email.

Two cautions before anyone wires this into a front desk. Google has not published API pricing yet, and the model is in public preview, so pilot it rather than build on it this week. And recording calls comes with consent rules that vary by state. Illinois requires everyone on the call to agree, so that step comes before any of the clever stuff.

If your business runs on phone calls and you want to know which of them are worth capturing first, that is exactly the kind of question our free process audit at New Face Design is built to answer.