Gemini 3.5 Transcribe

// Overview

Gemini 3.5 Transcribe is the speech-to-text pair Google announced on 26 August 2026: `gemini-3.5-transcribe` for pre-recorded audio and `gemini-3.5-transcribe-live` for real-time streaming. The Gemini API model list describes it as a "Low-latency speech-to-text model with utterance-based language detection, speaker diarization, and word timestamps." Google says the models automatically detect and transcribe over 85 languages.

The pitch is transcription that survives real speech rather than dictation. Google says the model "seamlessly handles self-corrections", "removes filler words" and "auto-formats text", identifies up to 3 speakers (more is experimental), returns word-level timestamps, and accepts a custom vocabulary for names and jargon. It also supports function calling, so a transcribed voice command can hand work to other Gemini models in the same flow.

On accuracy, Google reports an average word error rate of 4.0% for streaming and 2.6% for non-streaming use cases across noisy, real-world audio measured against Artificial Analysis benchmarks. On the multilingual FLEURS set across a group of top languages and locales it reports 5.50% WER streaming and 5.04% non-streaming. Against its predecessor Chirp 3, Google says time to final transcription improves by 70%.

Google launched the models in public preview in the Gemini API via Google AI Studio and Google Antigravity, and in public preview on the Gemini Enterprise Agent Platform. The same model powers the Gemini app on macOS (English), Rambler on Android in select countries and languages, with Chrome listed as coming soon. Google names Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents as launch partner platforms.

Pricing on the Gemini API is $2.00 per 1M input tokens (or $0.003 per minute of audio) and $12.00 per 1M output tokens (or $0.002 per minute of text) for the batch model. The live model is $3.50 per 1M input tokens (or $0.005 per audio minute) and $21.00 per 1M output tokens (or $0.004 per text minute). Both list a free-of-charge tier.

// Benchmarks

// Pricing

Rates for gemini-3.5-transcribe; Google also lists $0.003 per minute of audio input and $0.002 per minute of text output. The streaming model gemini-3.5-transcribe-live is $3.50 input / $21.00 per 1M tokens ($0.005 and $0.004 per minute). Both have a free-of-charge tier.

Pricing source ↗

// Strengths

• 5.50% word error rate on FLEURS (top locales) streaming — the lowest in Google's published launch comparison, ahead of Chirp 3 at 7.32%
• 2.6% average WER on pre-recorded audio and 4.0% streaming across noisy real-world audio, per Google's Artificial Analysis measurements
• Automatic language detection and transcription across more than 85 languages
• Speaker diarization for up to 3 speakers, word-level timestamps, and custom-vocabulary biasing for names and jargon
• Cleans up natural speech: handles self-corrections, removes filler words, auto-formats the transcript
• 70% faster time to final transcription than Chirp 3, by Google's measurement
• Function calling, so a transcribed command can trigger other Gemini capabilities in the same workflow

// Best for

• Reach for it for live captioning and voice agents, where the streaming endpoint runs through the Live API
• Reach for it for meeting and interview transcripts that need speaker labels and word-level timestamps
• Reach for it for multilingual audio where the language is not known in advance
• Reach for it for dictation flows that should read as clean prose rather than a literal transcript

// How to access