Google Launches Gemini 3.5 Transcribe Speech-to-Text Model

👁 113 reads

Google has released Gemini 3.5 Transcribe, a new speech-to-text model that edits out filler words and spoken corrections in real time to produce polished transcribed text. According to the company, the model is approximately 70 percent faster than its previous engine, Chirp 3, and cuts the live-speech error rate from 7.32 percent to 5.5 percent.

What Makes Gemini 3.5 Transcribe Different from Standard Speech Recognition

Traditional speech-to-text systems transcribe exactly what you say, including every hesitation, false start, and filler word. Gemini 3.5 Transcribe takes a different approach by understanding the intent behind your speech and cleaning it up automatically. When you say something like “I need to um, actually no, I want to schedule the meeting for Tuesday,” the model recognizes the correction and outputs clean text without the filler words or the false start.

Beyond raw accuracy, the model is designed to handle what Google calls the messiness of natural speech: removing “ums” and “uhs,” applying on-the-fly corrections, and referencing user-supplied custom vocabulary for specialized jargon. This means professionals in fields like medicine, law, or engineering can add their industry-specific terms, and the system will recognize and correctly transcribe them. It works across 85 languages and can process audio from up to three speakers in pre-recorded files, making it useful for transcribing meetings, interviews, and multi-person conversations.

Where You Can Use Gemini 3.5 Transcribe Today

The model already powers the Rambler feature inside Gboard on the Pixel 11. As of today it is also active in the Gemini app on macOS, in AI Studio’s build environment, and through the Gemini API for developers. Google’s Antigravity assistant is also gaining access, with permission-gated screen context and chat history. Support for Chrome, which would let users dictate into any web field, is described as coming soon, with Rambler set to expand to more devices later this year.

For everyday users, this means you can already start using the technology if you have a Pixel 11 smartphone or use the Gemini app on a Mac computer. Developers can integrate the transcription capabilities into their own applications through the API. The upcoming Chrome support will be particularly significant, as it will allow voice input across the entire web without needing special apps or extensions.

Performance Improvements Over Previous Technology

The speed and accuracy gains represent meaningful practical improvements. Being 70 percent faster than Chirp 3 means less waiting time between speaking and seeing your words appear on screen, which is crucial for real-time use cases like live captioning or voice commands. The reduction in error rate from 7.32 percent to 5.5 percent means fewer mistakes to correct manually, saving time and frustration especially during longer dictation sessions.

Frequently Asked Questions

How much more accurate is Gemini 3.5 Transcribe compared to the previous version?

Gemini 3.5 Transcribe reduces the live-speech error rate from 7.32 percent to 5.5 percent compared to Chirp 3, Google’s previous speech-to-text engine. It is also approximately 70 percent faster at processing speech.

What devices and platforms currently support Gemini 3.5 Transcribe?

The technology is currently available in the Rambler feature on Pixel 11 devices, the Gemini app on macOS, AI Studio’s build environment, through the Gemini API for developers, and in Google’s Antigravity assistant. Chrome support is coming soon, and Rambler will expand to more devices later this year.

How many languages and speakers does the model support?

Gemini 3.5 Transcribe works across 85 languages and can process audio from up to three speakers in pre-recorded files, making it suitable for multilingual environments and multi-person conversations.

// read moreWhisper AI: Speech-to-Text Transcription ReviewAI ToolsElevenLabs Tutorial: Text to Speech in MinutesGuidesAudio to Text Transcription: Free Tools & When They FailGuides

Kinetyra