Google just launched a new speech-to-text model called Gemini 3.5 Transcribe in public preview. Announced on August 26, 2026, on the Google Blog, it helps developers turn conversational audio into clean text.
Transcription often breaks on natural speech. This release aims to fix that.
Two ways to process audio
Google splits the model into two distinct tools:
- Live streaming audio. The Live API provides
gemini-3.5-transcribe-livefor bidirectional audio streams. - Recorded audio files. The Interactions API offers
gemini-3.5-transcribefor files, adding word timestamps and speaker labels. - Smart cleanup. The model removes filler words, fixes self-corrections, and handles custom terms and formatting.
- Broad language support. Google lists support for 85+ languages.
- Speaker detection. It identifies up to three distinct speakers, though support for 3+ speakers remains experimental.
You can test the preview in the Gemini API through Google AI Studio and Google Antigravity. Enterprise teams can access it through the Gemini Enterprise Agent Platform. Outside developer tools, the macOS Gemini app gets it in English, Rambler supports it on Android in selected countries and languages, and a Chrome rollout is coming soon.
What is not verified yet
Google and Artificial Analysis report strong benchmark numbers. They claim an average word error rate (WER) of 4.0% for streaming and 2.6% for recorded audio. They also report FLEURS WER scores of 5.50% live and 5.04% non-streaming, alongside a 70% speed improvement over Chirp 3.
Those numbers come directly from Google and Artificial Analysis tests. Nobody outside those teams has independently verified them yet. Real calls with heavy accents, background noise, or talking over each other may show different results. Google also did not share pricing, privacy terms, or data retention details yet.
How to test it yourself
Do not replace your existing speech system yet. Instead, run a direct comparison.
Gather sample files from your real workflow. Include noisy audio, tricky product names, speaker interruptions, and mixed languages. Run those test files through your current speech tool and the new Gemini model. Track accuracy, speed, speaker mistakes, and eventual cost.
If your meetings regularly have four or more people speaking, verify the speaker tracking carefully. Treat this launch as a promising preview to test, not a finished product ready for immediate deployment.