Gemini 3.5 Transcribe and Transcribe Live reach general availability

Google has taken its next-generation speech-to-text technology out of preview. In an August 26, 2026 update to the Gemini API changelog, Google announced that Gemini 3.5 Transcribe and Gemini 3.5 Transcribe Live are now generally available, giving developers two dedicated audio-understanding models built for production voice and transcription workloads.

What's new

• Two models, two workloads. gemini-3.5-transcribe handles non-streaming transcription; gemini-3.5-transcribe-live is a bidirectional streaming model delivered over WebSockets through the Live API.
• Non-streaming model: per the API changelog, it offers "high-accuracy, low-latency non-streaming speech-to-text with utterance-based language detection across 85+ languages, speaker diarization, word-level timestamps, and custom vocabulary biasing (up to 1,000 terms)."
• Streaming model: supports interim and finalized transcription events, a "Smart transcription mode," and multiple Voice Activity Detection (VAD) strategies for low-latency, always-on use cases.
• Accuracy claims. Google's product announcement describes the family as its "most precise speech-to-text model yet," citing an average Word Error Rate of 4.0% for streaming and 2.6% for non-streaming use cases, and the ability to attribute speech with timestamps for up to three speakers.
• Rollout surfaces. Beyond the raw API (via Google AI Studio), the underlying model is rolling out through Google Antigravity, the Gemini Enterprise Agent Platform, the Gemini app on macOS, and Rambler on Android in select countries and languages, with Chrome integration described as "coming soon."

Context

The release lands in the same API changelog that has carried a steady cadence of GA model launches through August, including Gemini 3.7 Flash on August 13. Rather than routing transcription through a general-purpose multimodal Gemini call, Google is now shipping dedicated, task-specific speech models — a shift that mirrors how OpenAI and others have separated out specialized endpoints (realtime audio, image generation) from their flagship chat models as usage matures beyond chatbot interfaces into agent and enterprise pipelines. Voice input has become one of the more contested surfaces in the assistant wars this year, with Google separately shipping Fn-key dictation inside its macOS Gemini app just a day earlier.

Why it matters

Speech-to-text is foundational infrastructure for voice agents, meeting transcription, accessibility tooling, and call-center automation — categories where latency and word-error rate translate directly into product quality and unit economics. By shipping a high-accuracy offline model and a low-latency streaming variant at the same time, and moving straight to general availability rather than an extended preview, Google is positioning Gemini's API as a default choice for developers who might otherwise stitch together specialized third-party ASR vendors. The custom vocabulary biasing and multi-speaker diarization features target exactly the enterprise use cases — meeting transcripts, support-call analytics — that have historically justified paying for dedicated speech vendors rather than a general-purpose LLM provider.