Google's Gemini 3.5 Transcribe cleans up your speech in real time, correcting yourself and cutting filler words automatically

[HighKey]

Google introduced Gemini 3.5 Transcribe this week, its most precise speech to text model yet, designed specifically to handle background noise, complex jargon, and natural speech disfluencies far better than conventional speech recognition systems typically manage. Rather than transcribing raw audio literally word for word, the model converts speech directly into accurate, polished, properly formatted text, automatically handling things like a speaker correcting themselves mid sentence or scattering filler words like um and ah throughout normal conversation.

The model ships across two separate APIs depending on the specific use case. A real time streaming version delivers continuous, bidirectional transcription with sub second latency for interactive voice applications through the Live API, while a separate version handles pre recorded audio, meetings, call logs, and similar recordings, adding speaker attribution and precise word level timestamps through the Interactions API. According to benchmarks from Artificial Analysis, the model achieves a word error rate of 4.0 percent in streaming mode and 2.6 percent for non streaming use, alongside a 70 percent improvement in time to final transcription compared to Google's previous transcription model, Chirp 3.

Beyond raw accuracy, the model brings several practical capabilities aimed squarely at real world dictation and voice interaction. It automatically detects and transcribes more than 85 languages while handling regional accents and dialects, supports custom vocabulary so it can correctly recognize specialized jargon or unusual spellings a user specifically provides, and can identify and attribute speech across up to three speakers in recorded audio, with experimental support for more. The model can also delegate more complex tasks like image generation or file analysis to other Gemini models through function calling, letting a spoken request trigger an entirely different kind of AI action automatically behind the scenes.

Google is already rolling the model out across multiple everyday surfaces, including a new Rambler feature on Android's Gboard keyboard that turns spoken thoughts into clean, formatted text while filtering out filler words, integration with Google Antigravity that uses screen context to improve transcription accuracy around technical file names and active documents, and voice powered features inside the Gemini app on macOS that pair natural speech with screen context to handle especially complex workflows. Developer platforms including Agora, LiveKit, and Pipecat have already built integrations on top of the Live API, and Google says early enterprise partners including Vivo and Intellitek Health have reported positive results specifically around the model's latency and broad language coverage

Intelligent transcription with Gemini 3.5 Transcribeblog.google

[John]

Curious how this specific model actually performs on the exact kind of speech accessibility challenges covered in that earlier EMMA AI receptionist story, accents, speech affected by conditions like a stroke, and other atypical speech patterns generally. Real inclusive accuracy across properly diverse speech patterns matters enormously more than raw benchmark numbers measured on standard, typical speech alone

[RandyOrton26]

85 plus languages with accent and dialect handling built in is actually impressive breadth if the accuracy numbers actually hold up consistently across all of them, rather than the benchmark figures being driven mainly by strong performance concentrated in just the handful of most common, well resourced languages

[HighKey15]

Google positioning this across so many different surfaces simultaneously, Gboard, Chrome, macOS, Antigravity, developer APIs, all at once suggests real internal confidence in the underlying model's actual quality and reliability. Companies generally don't roll something out this broadly, this fast, unless internal testing has already given them real confidence it's clearly ready for that kind of simultaneous, wide scale deployment

[MegaHavertz19]

Speaker attribution for up to three people, with more remaining experimental, feels like a meaningful current limitation worth flagging for anyone specifically hoping to use this for larger group meetings or multi person conference calls rather than smaller, more intimate conversations

[Joanne]

Automatically cleaning up self corrections and filler words during live transcription is a truly harder problem than it sounds on the surface. The model has to correctly figure out which earlier words to actually discard in real time as a sentence unfolds, rather than just transcribing everything literally and word for word as it comes in

[Keira85]

A 70 percent improvement in time to final transcription compared to the previous model is a substantial jump worth taking seriously rather than dismissing as a routine, incremental update. That kind of latency improvement is exactly what actually makes real time voice interaction feel really natural rather than noticeably laggy and awkward