Google introduced Gemini 3.5 Transcribe, a speech-to-text model designed to turn natural audio into accurate, clean, and well-structured text. It does more than recognize words: it also interprets corrections, removes filler words, identifies specialized vocabulary, and understands the context of a conversation.
What makes it different from conventional transcription? If you say, “See you Tuesday, no, Wednesday,” the model can deliver a corrected version that is ready to use.
Transcription designed for real conversations
Traditional speech recognition systems often struggle with background noise, accents, technical terms, or interrupted sentences. Gemini 3.5 Transcribe aims to address these obstacles by processing audio in a more intelligent way.
Its main features include:
- Automatic speech cleanup: removes filler words such as “um” and “uh,” while also organizing the text.
- Real-time corrections: recognizes self-corrections and preserves the speaker’s final intent.
- Custom vocabulary: lets you adapt the model to names, ZIP codes, order IDs, and technical terms.
- More than 85 languages: automatically detects languages and works with regional accents and diverse dialects.
- Speaker identification: for prerecorded audio, it can attribute segments to speakers and add timestamps for up to three participants. Support for more participants is still experimental.
This can be useful in many different scenarios, from a live-captioning app to the later analysis of customer service calls. It also opens the door to voice assistants that do more than transcribe: they understand what should happen next.
Two APIs for different use cases
Google offers Gemini 3.5 Transcribe through two interfaces within its developer platforms.
Real-time transcription
The Live API enables continuous two-way streaming with latency below one second through the gemini-3.5-transcribe-live model. It is designed for voice agents, instant captions, and interactive applications where waiting several seconds could disrupt the experience.
Processing prerecorded audio
The Interactions API uses gemini-3.5-transcribe to process meetings, calls, interviews, and other audio files. In this mode, it provides speaker attribution and word-level timestamps—important tools for conversation analysis and searches within recordings.
The practical difference is simple: one API follows the conversation as it happens, while the other analyzes the audio after it ends.
Accuracy, speed, and measurable results
According to measurements from Artificial Analysis, Gemini 3.5 Transcribe achieves an average word error rate, known as WER, of 4.0% in streaming and 2.6% in non-interactive scenarios. The lower this percentage, the fewer words are transcribed incorrectly.
The model also improves on Chirp 3, Google’s previous system. The company says that the time required to obtain the final transcription has been reduced by 70%.
In the multilingual FLEURS benchmark, Gemini 3.5 Transcribe recorded a WER of 5.50% in streaming and 5.04% in non-streaming processing across a set of major languages and regions. These results help show the scale of the improvement, although real-world performance can change depending on noise, microphone quality, language, and the number of voices present.
From dictation to AI-powered actions
Google’s approach does not stop at writing down what someone says. Gemini 3.5 Transcribe can serve as an entry point for carrying out tasks through voice, especially when combined with other Gemini models through function calling.
For example, in the Gemini app for macOS, someone can speak naturally, ask for a summary of local files, reuse text in another application, or request that an image be generated. The transcription model interprets the instruction and delegates complex tasks to other systems.
This evolution changes the idea of dictation. It is no longer just about converting speech into text, but about using voice as an interface for working with documents, applications, and digital agents.
Integration into Google products
Google is also incorporating the model into several of its own experiences:
- Gboard on Android: the Rambler feature turns spoken thoughts into formatted text, removes filler words, and lets you correct mistakes or change the style using your voice.
- Google Antigravity: with the user’s permission, it uses screen context and chat history to improve the transcription of file names, documents, and active processes.
- Google AI Studio: lets you use your voice in Build mode to create applications through spoken instructions.
- Gemini on macOS: combines transcription, screen context, and voice commands to automate workflows.
- Chrome: Google is preparing a feature for dictating text into any field on the web, from replies and posts to requests for Gemini.
The key is context. The same word can mean something different depending on the file you have open, the application you are using, or the previous conversation. A model capable of considering these signals can produce more useful results than an isolated transcriber.
Availability for developers and businesses
Gemini 3.5 Transcribe is available in public preview for developers through the Gemini API, Google AI Studio, and Google Antigravity. It is also coming to the Gemini Enterprise Agent Platform, while availability for Gemini Enterprise for Customer Experience is expected soon.
Platforms such as Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents are already working with the Live API to make it easier to build voice interfaces. These tools handle part of the streaming infrastructure, allowing teams to focus on designing the user experience.
For the general public, the model is available in the Gemini app for macOS in English and in Rambler for Android in certain countries and languages. Chrome integration will arrive later.
Gemini 3.5 Transcribe shows where interaction with AI is heading: speak naturally, correct yourself as you go, and expect the system to understand your full intent. The technology still depends on factors such as language, noise, and the privacy of the surrounding context, but the boundary between dictating an instruction and carrying out a task is becoming increasingly small.
Original source
https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5-transcribe