Introducing Gemini 3.5 Transcribe

What changed

Source summary

Google introduced Gemini 3.5 Transcribe, a new speech-to-text model offering real-time transcription with word error rates of 4.0% (streaming) and 2.6% (non-streaming), plus smart disfluency cleanup, custom vocabulary recognition and multi-speaker identification across more than 85 languages.

Why it matters to your work

Practical interpretation

Useful if you transcribe meetings, interviews or workshop recordings - multi-speaker identification and custom vocabulary support mean cleaner transcripts with less manual clean-up, across a wide range of languages.

When this could help

Worth evaluating now if you are building voice-dependent features or need better transcription accuracy, especially for noisy environments, specialised jargon, or multilingual use cases.

Gemini 3.5 Transcribe is Google's most precise speech-to-text model, with a 4.0% word error rate for streaming and 2.6% for non-streaming (Artificial Analysis benchmark). It handles self-corrections, removes filler words, auto-formats text, supports 85+ languages with auto-detection, offers speaker identification for up to 3 speakers, and delivers about 70% faster time-to-final-transcription than its predecessor Chirp 3.

How to get started

1. For real-time streaming, use model gemini-3.5-transcribe-live via the Live API. For pre-recorded audio with speaker attribution and timestamps, use gemini-3.5-transcribe via the Interactions API.

   Provider guidance ↗

Build it into your routine

SCALA suggested workflow

1. Voice agent accuracy test

   If you are evaluating speech-to-text providers, run a side-by-side comparison on your own audio samples (especially noisy or jargon-heavy ones) between your current provider and Gemini 3.5 Transcribe to see if the accuracy improvement applies to your use case.

What to check

• Speaker identification supports up to 3 speakers; beyond 3 is experimental. Consumer access is limited to English on macOS and select Rambler markets. Source ↗
• Function calling (handing off tasks to other Gemini models) is currently limited to the Gemini macOS app. Pricing is not stated. Source ↗