Google announced Gemini 3.5 Transcribe on August 26, 2026, a new speech-to-text model designed for real-time audio transcription and intelligent voice interactions. Rather than merely converting raw audio into words, the company says the model cleans up disfluencies and self-corrections, recognizes specialized terminology, and outputs formatted text that is closer to being ready for use.
The model targets developers building voice agents, real-time translation or captioning tools, and processing pipelines for analyzing calls and meetings after they have been recorded. It is available through the Gemini API in Google AI Studio, as well as through the Gemini Enterprise Agent Platform.
Two Interfaces for Different Use Cases
Google offers the model through two separate interfaces. The first is the Live API, using the gemini-3.5-transcribe-live model, which provides continuous bidirectional streaming with latency of less than one second, making it suitable for interactive voice conversation applications. The second is the Interactions API, using gemini-3.5-transcribe to process recordings, meetings, and call logs, with speaker attribution and word-level timestamps.
Google says Gemini 3.5 Transcribe can handle self-corrections, such as changing an appointment or a word while speaking, remove filler words such as stammering, and format text automatically. It also supports custom vocabularies that can be supplied to help it recognize specialized terminology and unusual spellings, and supports detecting and transcribing more than 85 languages while handling regional dialects and accents.
In recorded files, the model attributes speech to up to three speakers with timestamps, while support for more than three speakers remains experimental. It can also handle switching between languages during streaming, an important feature in multilingual conversations, but the source does not provide details about all supported languages or performance limits for each language.
Performance Figures Compared with Chirp 3
According to Artificial Analysis measurements cited by Google, the model achieved an average word error rate of 4.0% in streaming cases and 2.6% in non-streaming use cases. On the FLEURS benchmark across a range of languages and regions, it recorded a word error rate of 5.50% in streaming and 5.04% in non-streaming use. The company says latency to the final transcript improved by 70% compared with the previous Chirp 3 transcription model.
These figures indicate a dual improvement in accuracy and speed, but the comparison does not mean that results will be identical in every audio environment; they depend on the benchmark, language, recording quality, number of speakers, and noise level. The source also does not explain Artificial Analysis’s methodology in detail or provide a complete results table by language.
From Google Interfaces to Developer Tools
Google is already using the model in the Gemini app and in some Android experiments, including the Rambler feature, which converts speech into formatted text and allows filler words to be removed, corrections to be made, and the writing style to be changed by voice. In the Gemini app on macOS, voice commands can use screen context to perform tasks such as summarizing local files, reusing text between applications, or generating images by calling other Gemini models.
Google Antigravity uses screen context and conversation history, after obtaining the user’s permission, to improve recognition of active file and document names and agent ideas. The model can also be accessed in Build mode within Google AI Studio to build applications using voice. Google says the speak-to-type feature in any web field will soon arrive in Chrome.
What Changes in Practice?
The most important shift is the transition of audio transcription from an initial transcription step to a layer for understanding and cleaning speech and connecting it to the application’s context. For developers, this means the ability to build voice interfaces that do not require separate processing for every correction or filler word, with the option to use real-time streaming or archived recordings according to the nature of the product.
Gemini 3.5 Transcribe is currently available in public preview for developers through the Gemini API in Google AI Studio and Google Antigravity, and for enterprises through the Gemini Enterprise Agent Platform, with a later arrival planned for Gemini Enterprise for Customer Experience. For users, it is available in the Gemini app on macOS in English, while Rambler is available on Android in selected countries and languages, with an expected arrival in Chrome.