Gemini 3.5 Transcribe
2026-08-26 · Publisher : Google · Domain : Artificial Intelligence
Google launches Gemini 3.5 Transcribe, its most accurate speech recognition model, which turns audio into clean, formatted text, in streaming (Live API) or on recorded files.
What changes
- Handles self-corrections and removes filler words, with automatic formatting.
- Average word error rate of 4.0% in streaming and 2.6% non-streaming according to Artificial Analysis.
- Custom vocabulary, automatic detection of more than 85 languages.
- Speaker attribution for up to three speakers with word-level timestamps.
- Replaces Chirp 3: time to final transcript reduced by 70% according to Artificial Analysis.
### Specifications - Two models: gemini-3.5-transcribe-live (streaming via the Live API) and gemini-3.5-transcribe (files via the Interactions API). - Availability: Gemini API (AI Studio) and Gemini Enterprise Agent Platform. - Function calling to other Gemini models, available in the macOS app. ### Pricing Current pricing page: gemini-3.5-transcribe, input USD 2.00 per million tokens or USD 0.003 per minute, output USD 12.00 per million tokens; gemini-3.5-transcribe-live: input USD 3.50 per million tokens, output USD 21.00. Announcement-day pricing not verified. ### Benchmarks | Measure (scores published by the vendor) | Result | |---|---| | Average WER, streaming (Artificial Analysis) | 4.0% | | Average WER, non-streaming (Artificial Analysis) | 2.6% | | FLEURS WER, streaming | 5.50% | | FLEURS WER, non-streaming | 5.04% | ### Key takeaways A transcription building block for voice agents, captioning and call analysis, with configurable business vocabulary.
Links
Tags : Audio, Google, Gemini, Transcription, Voix