Intelligent transcription with Gemini 3.5 Transcribe

Gemini 3.5 Transcribe is in public preview for bidirectional streaming with sub-second latency and prerecorded audio with speaker attribution and word-level timestamps. Google says it detects more than 85 languages, supports custom vocabulary and live language switching, and attributes recorded speech to up to three speakers, with larger groups experimental.

Artificial Analysis measurements cited by Google report 4.0% average word error for streaming and 2.6% for non-streaming use. On FLEURS, Google reports 5.50% and 5.04%, plus a 70% reduction in time to final transcription over Chirp 3. The model removes fillers, resolves self-corrections, and formats text.

This is Google's launch post, so performance and selected customer praise need broader testing across accents, noise, specialist terms, and consequential identifiers. Average word error can obscure operational risk: one incorrect order ID may matter more than many correctly recognized common words.

The bigger shift is from faithful capture to editorial interpretation. Removing fillers and resolving corrections produces cleaner text but asks the model to decide which speech should survive. That helps dictation; legal, clinical, journalistic, and compliance uses may require raw audio plus a literal mode.

Signal Desk's read: Gemini 3.5 Transcribe is a voice application model, not merely improved speech-to-text. Function calls, screen context, chat history, and inline editing let speech trigger actions across software. Permission boundaries and confirmations therefore matter as much as recognition.

Distribution may be Google's advantage: APIs in AI Studio and its enterprise platform, the macOS Gemini app, Android Rambler, Gboard and Antigravity integrations, and planned Chrome availability. If quality holds, being present wherever people type may matter more than a narrow benchmark lead.

The two-API split also creates a clear product boundary. Streaming prioritizes continuous interaction, while prerecorded processing adds speaker attribution and word timestamps for meetings and calls. Teams should evaluate each mode separately because latency, cleanup, attribution, and audit requirements differ materially.

The Live API uses gemini-3.5-transcribe-live; the Interactions API uses gemini-3.5-transcribe for meetings, calls, and recordings. Several developer platforms support related real-time media infrastructure.

In the macOS app, functions can delegate file analysis and image generation to other Gemini models. Antigravity can use screen context and chat history with permission to recognize filenames and active documents.