Google's new Gemini 3.5 Transcribe model cleans up voice input across products
Google has introduced Gemini 3.5 Transcribe, a speech-to-text model that filters out filler words and self-corrections to produce polished text. The model is about 70 percent faster than its predecessor and has a lower error rate, supporting 85 languages and up to three speakers in pre-recorded audio. It already powers the Rambler feature on Pixel 11 and will be integrated into more Google products, including Chrome.
The model builds on Google’s prior Chirp 3 engine, cutting latency by roughly 70% and lowering live-speech errors from 7.32% to 5.5%. It supports 85 languages and can distinguish up to three speakers in recorded audio, while also referencing user-defined vocabularies for technical terms. Beyond Gboard’s Rambler on Pixel 11, it now appears in the Gemini macOS app, Antigravity, AI Studio, and the Gemini API. Chrome integration is planned, which would let users dictate into any web form. The system actively rewrites spoken text, removing fillers and self-corrections, so output reflects intended meaning rather than exact words.
This tool could reshape how people interact with digital devices, making voice input more practical for drafting emails, coding, or chatting. By cleaning speech automatically, it may reduce editing time but also raises questions about fidelity—users might lose control over exact wording. Broader adoption across Chrome and Pixel devices could normalize AI-mediated communication, potentially affecting how individuals trust or verify transcribed content in professional or legal contexts.