Google Rolls Out Gemini 3.5 Transcribe with Exceptional Accuracy

Google has just announced the launch of Gemini 3.5 Transcribe, a speech-to-text model designed to handle messy human speech in real time, across Android, Mac, and web apps.

Instead of simply turning spoken audio into literal text, the model filters out background noise, removes filler words like “um” and “ah,” and cleans up accidental speech mistakes as people talk.

Traditional dictation systems often fall apart when people hesitate or change their minds mid-sentence. If a user says, “Let’s meet Tuesday… actually, no, Wednesday,” classic speech software writes down the entire mistake. Gemini 3.5 Transcribe processes the intent behind the words. It drops the self-correction automatically and outputs a clean final sentence reading “Let’s meet Wednesday.”

The software also strips away awkward pauses and handles complex terms, numbers, and technical jargon without freezing up. Benchmark tests from Artificial Analysis show the streaming version of the software hits a 4.0% Word Error Rate, while the pre-recorded audio option hits 2.6%.

Developers can access the model starting today in public preview through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. The integration comes in two variations.

The first is a Live API (gemini-3.5-transcribe-live) built for continuous, bidirectional streaming with sub-second latency. This version powers fast-paced interactive voice setups, real-time captioning tools, and quick dictation. The second option is an Interactions API (gemini-3.5-transcribe) built for processing long pre-recorded media. It includes automated speaker identification for up to three people, along with precise word-level timestamps for call logs and meeting notes.

Google is also weaving the technology directly into everyday software. On Android devices, the tech powers a feature called Rambler inside Gboard. Users can talk out loud to generate formatted text, edit misspellings, or swap writing styles using quick voice prompts.