Skip to content
See the World Through Science
News

Google's New Transcription Model Removes the 'Ums' Before You See Them

AI & Technology

Republish this story

Our work is licensed under Creative Commons BY-NC 4.0. You may republish this piece for free — with credit to ALLATRA Media and a link to the original, unedited beyond length trims, and not for commercial use.

Read the full license

An early Dictaphone wax-cylinder dictating machine with a speaking horn and flexible tube, photographed in a museum display.
An early Dictaphone dictating machine, which recorded speech for a typist to transcribe by hand. A historical illustration of the task Google's model now does automatically; the machine has no connection to the release."File:Edmonds Historical Museum - Dictaphone 03.jpg" by Joe Mabel, via wikimedia, CC-BY-SA-3.0 · CC-BY-SA-3.0

Google has released Gemini 3.5 Transcribe, a speech-to-text model, into public preview, the company announced Aug. 26. It is already running in two consumer products: the Gemini app on macOS, in English, and the Rambler dictation feature in Gboard on Android, in selected countries and languages.

Developers reach the model through the Gemini API in Google AI Studio and in Google Antigravity; businesses through the Gemini Enterprise Agent Platform. Google splits it across two interfaces. One, gemini-3.5-transcribe-live, runs on the company's Live API for continuous two-way streaming with what Google describes as sub-second latency. The other, gemini-3.5-transcribe, runs on its Interactions API for recorded audio and adds speaker labels and word-level timestamps.

On figures Google attributes to Artificial Analysis, the model averages a 4.0% word error rate, the share of words it transcribes wrongly, in streaming use, and 2.6% on pre-recorded audio. Google also attributes to Artificial Analysis a 70% improvement in the time taken to reach a final transcript, measured against Chirp 3, its previous transcription model.

A second set of figures is Google's own and carries no outside attribution. On the FLEURS multilingual benchmark, "across a set of top languages and locales," the company reports word error rates of 5.50% in streaming mode and 5.04% for pre-recorded audio, both an improvement on Chirp 3.

The announcement compares the model only with Chirp 3. It gives no head-to-head results against competing products, and there is no independent evaluation of Gemini 3.5 Transcribe yet.

Google says the model detects and transcribes more than 85 languages automatically, attributes speech to as many as three speakers in recorded audio, with support beyond three marked experimental, and strips filler words, resolves self-corrections and formats text as it goes.

Dictation in Chrome, and a version for Gemini Enterprise for Customer Experience, are listed as coming soon.

Sources

By Oli KotykWriter

Spot an error?

Spot an error?

Report an error

Spotted a mistake on this page? Tell us what's wrong and our editors will take a look.

What kind of problem?

Only if you'd like us to be able to follow up. We won't use it for anything else.

We correct mistakes openly. Select any text to flag it. Fixes are logged under our Corrections Policy.

Report an error

Reporting on

Google's New Transcription Model Removes the 'Ums' Before You See Them

What kind of problem?

Only if you'd like us to be able to follow up. We won't use it for anything else.

We read every report. Corrections are logged publicly.