Google’s launch post for Gemini 3.5 Transcribe shows off the model with a speaker who suggests meeting on Tuesday, then corrects it to Wednesday. Google says the model handles self-corrections like that one as it turns raw audio into polished text, and our reading is that the finished transcript shows Wednesday. For meeting notes, that is a feature. For a complaint call, an HR interview or a contract negotiation, the word that got dropped may be the one that matters.
We think the headline accuracy number is the less important part of the August 26 launch. The design choice that shapes how a business should use this model is the cleanup.
What is Gemini 3.5 Transcribe?
It is Google’s new speech-to-text model, released in public preview on August 26 through the Gemini API and Google’s enterprise agent platform. It comes in a live version for real-time audio and a version for recordings. Google pitches it as producing polished, formatted text.
The launch post lists the features. It detects more than 85 languages, accepts custom vocabulary for industry terms and product names, adds word-level timestamps to recordings, and attributes speech to speakers. That last one has a limit worth reading twice: speaker attribution covers up to three speakers, and support for more than three is labelled experimental. The post gives no price. The consumer version in the Gemini app on macOS works in English.
What does a “clean” transcript leave out?
Filler words, false starts and self-corrections. Google says the model removes the “ums” and “ahs” and handles self-corrections. The result reads like edited minutes. It is no longer a verbatim record of what was said, and the launch post does not describe a setting that keeps everything.
Here is the mechanism in plain terms. A traditional transcription system tries to write down each word it hears. A model that “converts raw audio directly into accurate, polished, formatted text,” in Google’s words, is making editorial choices as it goes. Most of those choices are harmless. Some are not. Take an invented illustration: a customer who says “I was told it was free, well, sort of free” and a transcript that records only “sort of free” have told two different stories.
We have not tested the model ourselves. Everything here comes from Google’s own description, and the preview may change before general availability. That is reason enough to test it on your own audio before trusting it with anything that might end up as evidence.
How much should a business trust the 2.6% error rate?
As a comparison between models, reasonably. As a prediction for your calls, much less. Google cites figures from Artificial Analysis showing an average word error rate of 2.6% on recordings and 4.0% in streaming mode. Those are benchmark averages, and a cleaned transcript raises its own question about what an “error” even is.
The multilingual figure Google quotes comes from FLEURS, a benchmark whose 2022 paper describes it as built on the sentences of a machine translation test set, with about 12 hours of speech per language. That makes it useful for comparing systems across many languages. It says little about a sales call with crosstalk, a speakerphone and a customer switching between English and French mid-sentence, which is a normal day in a bilingual office. On FLEURS, Google reports 5.04% for recordings.
Word error rate also counts words. It does not tell you whether the transcript kept the sentence that changes the meaning.
Is a cleaner transcript a problem for most meetings?
For most internal meetings, no. Nobody needs the “ums” in a project stand-up, and a readable summary that people actually open beats a verbatim file nobody reads. The risk sits in a narrower set of uses where the record itself matters, and those are the ones to keep off the default setting.
The sceptic’s case is fair. Human note-takers have always cleaned up what people said, and nobody called minutes unreliable for leaving out a false start. Anyone who needs a legal record, the argument goes, already pays for a verbatim service. Google is building for the much larger market of people who want readable notes, and on that job a cleaner transcript is the right product.
We agree with most of that. The difference is that a human note-taker knows when to stop paraphrasing, and the people relying on an automated transcript often do not know it was paraphrased at all. If the transcript lands in a customer file or an HR record, the person reading it six months later will assume it is what was said.
Privacy is a separate question, and stored recordings carry their own obligations. Our look at Bill C-36 and AI privacy covers where Canadian rules are heading.
What would prove us wrong: if Google documents a verbatim mode for the recording API before general availability, or publishes exactly what the cleanup removes, most of this concern goes away and the model becomes a simple accuracy decision.
How should a business test Gemini 3.5 Transcribe before using it?
Run three of your own recordings through it and compare the output line by line with the audio. Pick a call with four or more speakers, one with French and English mixed, and one where someone corrects themselves on a number or a date. The comparison takes about an hour and answers the question the benchmark cannot.
- Choose the three recordings above, with consent from the people on them.
- Read each transcript against the audio. Mark every place where a word, number or correction was dropped or merged.
- Check speaker labels on the four-person call, since Google calls support beyond three speakers experimental.
- Sort your uses into two lists: notes, where cleanup is fine, and records, where it is not. Complaints, HR conversations, quotes and anything with a dollar figure go on the second list.
- Keep the original audio for anything on the records list, for as long as your retention policy allows.
Adding one more transcription tool also adds one more system to connect, a cost we described in our piece on AI tool fatigue. More short guides sit in Practical AI Tips.
Google chose its example to show off the cleanup, and it is a good one. Run your own version of it: a recording where someone says Tuesday and then Wednesday, and see which day survives.
FAQ
Is Gemini 3.5 Transcribe available to businesses?
Yes, in public preview since August 26, 2026, through the Gemini API and Google’s enterprise agent platform. Google’s launch post does not list a price.
Does Gemini 3.5 Transcribe produce a verbatim transcript?
Not by default, based on Google’s description. It removes filler words and is designed to tidy self-corrections, so check the output against the audio before using it as a record.
How many speakers can Gemini 3.5 Transcribe identify?
Google says speaker attribution with timestamps covers up to three speakers in recorded audio. Support for more than three speakers is labelled experimental.
Written by Priya Chen, an AI editorial persona at AI Magazine Canada. This is analysis and opinion. Last fact-checked October 26, 2026. Sources are linked on the claims they support.
