Gemini 3.5 Transcribe
The most precise speech-to-text model yet
Context window
96000
Input / 1M tokens
2.50
Output / 1M tokens
12.00
Provider
Data verified 2026-09-03
<cite index="6-2">Google's most precise speech-to-text model, designed for intelligent voice interactions.</cite> <cite index="6-1">Unlike conventional speech recognition models that struggle with background noise, complex jargon, and disfluency cleanup, Gemini 3.5 Transcribe converts raw audio directly into accurate, polished, formatted text.</cite>
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Available in public preview via Gemini API in Google AI Studio and Gemini Enterprise Agent Platform. Also available in Gemini app on macOS and Gboard on Android.
Strengths
- ✓Supports 85+ languages with accent and dialect handling
- ✓Removes filler words and handles self-corrections automatically
- ✓Speaker attribution and word-level timestamps for up to 3 speakers
- ✓Custom vocabulary recognition
- ✓Function calling for delegating tasks to other Gemini models
Best for developers who...
When to choose it (and when not to)
Reach for Gemini 3.5 Transcribe when...
- →When you need accurate transcription with speaker attribution
- →For multilingual applications requiring 85+ language support
- →When building voice agents with function calling capabilities
Look elsewhere if...
- ✕If you need a general-purpose LLM rather than a specialized speech-to-text model
- ✕For use cases requiring offline processing
How to use it
Quickstart
Pythonimport anthropic
# Use Gemini API for transcription via Google AI Studio or Enterprise PlatformAccessible through Gemini API in Google AI Studio and Gemini Enterprise Agent Platform
API model id: gemini-3.5-transcribe
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| FLEURS (multilingual) | 5.50% WER (streaming), 5.04% WER (non-streaming) | Across top languages and locales, improving over Chirp 3 |
| Artificial Analysis measurement | 4.0% WER (streaming), 2.6% WER (non-streaming) | On diverse real-world conditions including background noise |
| Word Error Rate (WER) - Non-streaming | 2.6% | As measured by Artificial Analysis |
| Word Error Rate (WER) - Streaming | 4.0% | As measured by Artificial Analysis |
| FLEURS multilingual benchmark - Streaming | 5.50% | Google reported on top languages/locales |
| FLEURS multilingual benchmark - Non-streaming | 5.04% | Google reported on top languages/locales |
| Word Error Rate (streaming) | 4.0% | Artificial Analysis measurement |
| Word Error Rate (non-streaming) | 2.6% | Artificial Analysis measurement for English |
| FLEURS multilingual WER (streaming) | 5.50% | Google reported measurement |
| Time to Final Transcription | 70% improvement | vs Chirp 3; streaming variant: 0.40s latency from end of speech |
Source: Google Official Announcement
Compare Gemini 3.5 Transcribe
Compare Gemini 3.5 Transcribe with any other model
Build a comparison →All model comparisons →