Gemini 3.5 Transcribe: What Builders Need to Know
Google's new speech model returns what you meant, not what you said. That is a gift for meeting notes and a problem for anything that needs a verbatim record.
Gemini 3.5 Transcribe is Google's new speech-to-text model, announced on 26 August 2026 and available in public preview through the Gemini API. It ships as two endpoints rather than one, reports a 2.6% average word error rate on pre-recorded audio and 4.0% on live streams, and covers more than 85 languages with automatic detection. The interesting part is not the accuracy number. It is that the model returns cleaned-up text rather than a literal transcript, which changes what you have to build around it.
What Gemini 3.5 Transcribe actually returns
Most speech-to-text models give you what was said. This one gives you what was meant.
Per Google's announcement, the model drops filler words, resolves self-corrections, and formats the output. If somebody says "let's meet Tuesday, no, Wednesday," the transcript says Wednesday. The ums and ahs are gone before you see them.
That is a real convenience and a real trap, depending on what you are building.
If you are building | Cleaned output is | Because |
|---|---|---|
Meeting notes, summaries, dictation | A gift | You were going to strip fillers anyway, and now you skip a whole cleanup pass |
Support call analysis | Mostly good | Sentiment and intent survive; hesitation cues do not |
Legal, medical, or compliance records | A problem | A record of what was said cannot silently become a record of what was meant |
Speech research, accessibility captions | A problem | Disfluencies are data, not noise |
Nobody is going to warn you about that fourth row in a launch post. If your use case needs a verbatim record, the cleanup is a feature you have to work around rather than one you get for free.
The two endpoints are not interchangeable
The model comes as a pair. `gemini-3.5-transcribe` handles pre-recorded audio through the Interactions API. `gemini-3.5-transcribe-live` handles live audio through the Live API.
The split matters because the accuracy figures differ, and they differ in the direction you would expect. Google reports 2.6% average WER for the non-streaming case and 4.0% for streaming. On the public FLEURS benchmark the two land at 5.04% and 5.50%.
That gap is the cost of not being able to look ahead. A streaming model has to commit to a word before it has heard the rest of the sentence. A batch model gets the whole recording and can revise. If your product can tolerate a few seconds of delay, batch is the more accurate choice and you should not reach for the live endpoint out of habit.
Google also claims time to final transcription improves by 70% against Chirp 3, its previous model. That is a latency claim about the same-shaped work, not a claim about accuracy.
How to read a 2.6% word error rate
A word error rate is the share of words the model got wrong, counting insertions, deletions and substitutions against a reference transcript. We wrote a fuller explainer on what word error rate measures and where it misleads, because the number is easy to over-trust.
Two things to hold on to here. First, 2.6% is an average across 85+ languages, and averages across languages hide enormous spread; the language you care about may be well above or well below it. Second, the FLEURS figure of 5.04% is roughly double the headline average, which tells you the headline is measured on something friendlier than a standard public benchmark. Neither of those makes the model bad. Both mean the number you should plan against is the one you measure on your own audio.
Where it is available right now
Preview, not general availability. Developers reach it through the Gemini API via Google AI Studio and Google Antigravity. Enterprises get it through the Gemini Enterprise Agent Platform. On the consumer side it already powers Rambler on Android in selected countries and the Gemini app on macOS in English, with Chrome dictation described as coming.
Preview status is the practical constraint. Preview endpoints change, and a model that reformats its output is a model whose output shape can change under you. If you ship on this, pin your expectations in tests rather than in your assumptions.
If you want to try speech input in something you have built, we walked through the plumbing in adding voice input to an AI-built app, which is provider-agnostic and covers the parts nobody warns you about.
Should you switch?
If you are already on a speech-to-text provider and it works, a preview model with a better benchmark is not a reason to migrate. If you are choosing now, or your current transcripts need a cleanup pass you wrote yourself, this is worth an afternoon of testing on your own audio.
The general habit here matters more than this specific model. Every fortnight brings a release with a better number attached, and the discipline of not chasing every model release is worth more than any individual switch.
FAQ
Is Gemini 3.5 Transcribe free?
Google's announcement does not state pricing. It is in public preview through the Gemini API, and preview pricing is usually published on the API pricing page rather than in the launch post, so check there before you plan a budget around it.
What is the difference between the two Gemini 3.5 Transcribe endpoints?
`gemini-3.5-transcribe` processes pre-recorded audio through the Interactions API and reports 2.6% average WER. `gemini-3.5-transcribe-live` handles real-time audio through the Live API and reports 4.0%. Use the batch endpoint whenever your product can wait.
Does it keep filler words if I want them?
The announcement describes filler removal and self-correction resolution as what the model does, not as a toggle. If you need verbatim output, test whether prompting or configuration can turn it off before committing, and assume it cannot until you have proven otherwise.
How many languages does it support?
More than 85, with automatic detection, so you do not have to declare the language up front. Per-language accuracy is not broken out in the announcement, which matters if your users are not speaking English.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
