Gemini 3.5 Transcribe Reaches General Availability
Gemini CLIGoogle has taken Gemini 3.5 Transcribe generally available in the Gemini API, shipping two dedicated speech-to-text models built on Gemini's audio understanding. Gemini 3.5 Transcribe (gemini-3.5-transcribe) handles high-accuracy, non-streaming transcription with speaker diarization and word-level timestamps, while Gemini 3.5 Transcribe Live (gemini-3.5-transcribe-live) adds low-latency, bidirectional streaming over WebSockets via the Live API. Independent benchmarks and coverage cite word-error rates around 4.0% for streaming and 2.6% for non-streaming audio, positioning it as Google's most precise speech-to-text model yet. The models support 85+ languages, custom vocabulary biasing, and are already integrated by third-party platforms including LiveKit, Pipecat, and Vercel.
Key Takeaways
- Google shipped two distinct models — a non-streaming
gemini-3.5-transcribeand a real-timegemini-3.5-transcribe-live— rather than a single one-size-fits-all transcription endpoint. - Reported word-error rates of 2.6% (non-streaming) and 4.0% (streaming) put the models in competitive territory with dedicated voice-AI providers, according to independent coverage.
- Speaker diarization and word-level timestamps make the non-streaming model well-suited for meeting transcripts, call-center logs, and captioning workflows.
- Smart transcription mode automatically strips filler words and self-corrections and applies punctuation/formatting, reducing the need for manual post-processing.
- Custom vocabulary biasing supports up to 1,000 domain-specific terms, letting teams tune recognition for jargon-heavy or proper-noun-heavy audio.
- Third-party platforms including LiveKit, Pipecat, Vercel, and LangChain have already integrated the Live model, signaling early ecosystem support beyond Google's own apps.
Sources & Mentions
4 external resources covering this update
Google launches Gemini 3.5 Transcribe, which powers Gboard Rambler & is coming to Chrome
9to5Google
Google rolls out Gemini 3.5 Transcribe to improve real-time dialogue and speech recognition
Android Authority
Google says its latest Gemini transcription model can turn your ramblings into structured text
Engadget
Google's latest speech-to-text Gemini model offers a platter of new features
Neowin
Two New Speech-to-Text Models
Google has moved Gemini 3.5 Transcribe into general availability within the Gemini API, introducing two purpose-built models for converting audio into text. The standard Gemini 3.5 Transcribe model (gemini-3.5-transcribe) is designed for high-accuracy, low-latency non-streaming transcription, while Gemini 3.5 Transcribe Live (gemini-3.5-transcribe-live) targets real-time, bidirectional streaming use cases over WebSockets through the Live API. Both models are built on Gemini's broader audio-understanding capabilities rather than being adapted from a legacy speech-recognition stack, which Google says gives them stronger context awareness than conventional transcription systems.
Accuracy and Language Coverage
Google is positioning this as its most precise speech-to-text model to date. Reported word-error rates land around 2.6% for pre-recorded, non-streaming audio and 4.0% for streaming audio, benchmarks that independent outlets have highlighted as competitive with dedicated voice-AI vendors. The models perform automatic, utterance-based language detection across more than 85 languages and are built to handle regional accents and dialect variation without requiring the caller to specify a language up front.
Developer-Facing Features
For developers, the headline additions are speaker diarization and word-level timestamps, which let an application attribute transcribed segments to individual speakers and pinpoint the exact timing of each word — useful for meeting transcripts, call logs, and captioning pipelines. A Smart transcription mode automatically cleans up disfluencies, filler words, and self-corrections, and applies formatting such as capitalization, punctuation, and inverse text normalization so the output reads as polished prose rather than a raw transcript. Custom vocabulary biasing lets teams pass up to 1,000 domain-specific terms, acronyms, or proper names to improve recognition accuracy for specialized content.
Real-Time Streaming via the Live API
Gemini 3.5 Transcribe Live is aimed at interactive voice applications that need sub-second latency. It supports both interim and finalized transcription events and offers multiple Voice Activity Detection (VAD) strategies, giving developers control over how aggressively the model segments speech during a live session. Several third-party platforms — including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents — have already integrated the model through the Gemini Live API.
Availability
The models are accessible to developers through Google AI Studio and Google Antigravity as part of the Gemini API. Pre-recorded transcription is available through the model endpoint directly, while continuous streaming runs through the Live API. The underlying transcription technology is also rolling out in first-party Google products, including the Gemini app on macOS, the Rambler feature on Android, and — according to reporting — is coming soon to Chrome.