Google Launches Gemini 3.5 Transcribe for Low-Latency Speech-to-Text

What Google released

On August 26, 2026, Google introduced Gemini 3.5 Transcribe, a dedicated speech-to-text model for intelligent voice applications. Google lists two developer-facing variants: gemini-3.5-transcribe-live for streaming transcription through the Live API and gemini-3.5-transcribe for prerecorded audio through the Interactions API. The model is in preview rather than general availability.

Core capabilities

Google says the model automatically detects and transcribes more than 85 languages, including language switching within a session. It supports custom vocabulary biasing, smart transcription that removes filler words and resolves spoken self-corrections, and word-level timestamps for prerecorded files. Speaker diarization is supported for recorded audio, with support beyond three speakers described as experimental.

Real-time and file workflows

The live variant is designed for low-latency streaming voice interfaces, captioning and conversational systems. The prerecorded variant supports up to one hour of audio per request, with shorter limits when diarization or detailed timestamps are enabled. Google positions the model for voice agents, real-time captions, meeting and call transcription, post-call analytics and domain-specific dictation.

Why it matters

Speech-to-text is becoming a core input layer for AI agents. A specialized Gemini transcription model separates fast speech recognition from broader multimodal reasoning, which can help developers build voice products with more predictable latency and transcription controls. Google also exposes the model through Google AI Studio and its Gemini developer platform, making it accessible without building a separate speech stack.

Released versus upcoming

Gemini 3.5 Transcribe is available in public preview now. Developers should not confuse this with unrelated unreleased Gemini 3.5 variants; production teams should continue to treat preview APIs, quotas and behavior as subject to change until Google announces general availability.