Google on Wednesday introduced Gemini 3.5 Transcribe, a speech-to-text model designed to enhance real-time and batch transcription capabilities across enterprise and consumer applications.
The model delivers a 2.6% word error rate in non-streaming applications and 4.0% in streaming mode, according to Artificial Analysis, outperforming the prior Chirp 3 model. Latency improvements are substantial, with a 70% reduction in time to final transcription and sub-second real-time streaming latency. On the FLEURS benchmark, the model recorded word error rates of 5.04% for non-streaming and 5.50% for streaming use cases.
Gemini 3.5 Transcribe supports automatic detection and transcription of more than 85 languages and can identify up to three speakers in pre-recorded audio with timestamps. It processes audio into formatted text, handles background noise and technical terminology, and removes filler words while allowing speaker corrections and custom vocabulary recognition.
The model is accessible via two APIs: the Live API for real-time streaming using the `gemini-3.5-transcribe-live` endpoint, and the Interactions API for pre-recorded audio processing through the `gemini-3.5-transcribe` endpoint, which includes speaker attribution for meetings and call logs.
Integration is underway across Google’s ecosystem, including the Gemini app on macOS, the Rambler feature on Android via Gboard, and Google AI Studio. Developers can access the model in public preview through Google AI Studio, while enterprises can deploy it via the Gemini Enterprise Agent Platform. A dictation feature is planned for Chrome in a future update.
