Google has unveiled Gemini 3.5 Transcribe, its new AI model specializing in converting speech to text. The company describes it as its most accurate system to date, designed for intelligent voice interactions. Unlike traditional systems, this model not only transcribes what is said, but also interprets the speaker's intent, corrects errors on the fly, and produces clean, formatted text.
The model is designed for both real-time and recorded use, and promises to perform well in noisy environments, with strong accents, or with technical jargon. Furthermore, it automatically detects over 85 languages, recognizes up to three speakers, and adds word-by-word timestamps. All of this is achieved with a remarkably low error rate, according to independent testing.
Intelligent transcription: more than just dictation
One of the key features of Gemini 3.5 Transcribe is its ability to understand self-corrections. For example, if the speaker says, "Let's meet on Tuesday... no, on Wednesday," the model will use the second option. It also eliminates filler words like "uh" or "um" and automatically applies punctuation and formatting. Furthermore, it allows you to add custom vocabulary to recognize proper nouns, technical terms, or uncommon words. According to data published by Google, the error rate per word is 4% for streaming audio and 2,6% for pre-recorded audio, as measured by Artificial Analysis.
Another standout feature is speed. Compared to its predecessor, Chirp 3, Google claims that the time to obtain the final text is reduced by 70%. In live streaming, latency is less than one second, enabling its use for real-time captioning or voice assistants.
Integration into the Google ecosystem
Gemini 3.5 Transcribe is already integrated into several Google products. On Android, Gboard's Rambler feature uses this model for voice-to-text input. It's also available in the Gemini app for macOS, allowing users to use voice commands with screen context to summarize files, generate images, or reuse text. Additionally, Google Antigravity incorporates it to combine screen context and chat history.
The company has also announced that the feature will soon be available in Chrome. This will allow users to dictate text in any web field, such as emails, forms, or posts, without typing. While there is no specific date, this feature is expected to expand the use of voice transcription in the browser.
Availability for developers
For developers, Gemini 3.5 Transcribe is available in public preview through the Gemini API, in Google AI Studio, and in the Gemini Enterprise Agent Platform. Two versions are offered: the Live API for real-time, bidirectional streaming with sub-one-second latency, and the Interactions API for processing pre-recorded audio with speaker attribution and timestamps. The former is designed for voice agents and live captioning, while the latter is useful for meetings, interviews, or call analytics.
The model can also delegate complex tasks to other Gemini models, such as image generation or file analysis, through function calls. This capability is currently available in the Gemini app for macOS.
Google's commitment to voice as the primary interface is reinforced with Gemini 3.5 Transcribe. Its combination of accuracy, speed, and intelligent features makes it a key tool for both users and developers. With its imminent arrival in Chrome, voice dictation could become commonplace on the web. It remains to be seen how it will adapt to more demanding use cases, but initial data is promising.