Google's Gemini 3.5 Transcribe targets voice agents and call notes
On 26 August, Google released Gemini 3.5 Transcribe, its new speech-to-text model, in public preview for developers. Speech-to-text is the step that turns what someone says on a call or into a microphone into written text, and it is the first link in almost every voice agent and every tool that writes up meeting notes.
Google says the model works from raw audio and returns formatted text, removing filler words like um and ah and handling a speaker who corrects themselves mid-sentence.
Two versions for two jobs
There are two ways to call it. The live versiongemini-3.5-transcribe-livestreams text back in under a second through Google's Live API and is meant for voice agents and live captions. The batch versiongemini-3.5-transcribeprocesses recordings such as meetings and call logs, with word-level timestamps and labels for who spoke. Speaker labels work for up to three people, and Google calls support for more than three experimental.
Google also says it detects and transcribes more than 85 languages, accepts a custom vocabulary list for jargon and unusual spellings, and handles details like postal codes and order IDs in noisy audio.
The numbers Google gives
Measured by Artificial Analysis, an independent benchmarking company, the average word error rate is 4.0% for streaming and 2.6% for recorded audio. Word error rate is the share of words the model gets wrong, so lower is better. Against Chirp 3, Google's previous speech model, Google says time to a final transcript improved by 70%. The post does not include pricing.
The same model already runs Rambler, a dictation feature in Gboard on Android, and voice input in the Gemini app on macOS. Google says talk-to-type in any web field is coming to Chrome.
What to do if you run a voice line
If you have a voice agent or a call-summary tool, the useful test is your own audio. Take ten real recordings with background noise, names and order numbers, run them through the batch version in Google AI Studio and count the mistakes on the words that matter to you, such as names, dates and amounts. Fill the custom vocabulary with your product names before you start. Keep in mind it is a preview, so check the terms before you move live calls to it.
What we think
For anyone building phone agents, transcription errors cause more trouble than people expect, because a wrong digit in an order number sends the rest of the conversation the wrong way. A 4.0% average error rate on a third-party benchmark is a good sign. The three-speaker limit on labels matters less for phone calls, which usually have two voices, and more for meetings.
Sources