Google Expands Voice-to-Text Capabilities with New Gemini 3.5 Transcribe Model

While tech enthusiasts continue to speculate on the arrival of the highly anticipated Gemini 3.5 Pro, Google is moving forward with the release of a specialized model within the 3.5 family. The company has officially unveiled Gemini 3.5 Transcribe, an advanced AI solution designed to revolutionize voice input by transforming raw, messy speech into polished, coherent text.

Beyond Simple Transcription

Previously available only as the engine behind the “Rambler” feature on the Pixel 11, Gemini 3.5 Transcribe is set to roll out across the broader Google ecosystem. The model goes significantly beyond traditional speech-to-text engines by actively interpreting user intent.

As users dictate, the AI performs real-time cleanup, intelligently removing “ums,” “uhs,” and other fillers that often clutter spoken sentences. It also handles on-the-fly self-corrections, ensuring the final output reads as if it were written rather than spoken. For professional or technical users, the system supports custom vocabulary to accurately capture specialized jargon, providing a more reliable experience for niche industries.

Speed and Precision

Google claims that the new model offers a substantial performance upgrade over its predecessor, Chirp 3. In terms of latency, the new model is reportedly 70 percent faster from the moment of speech to the delivery of the final text.

The accuracy metrics have also seen a refined improvement. Google reports that the live-speech error rate for Gemini 3.5 Transcribe sits at 5.5 percent, a notable step up from Chirp 3’s 7.32 percent error rate. While the incremental decrease in error percentage might seem modest, even small improvements in transcription accuracy significantly reduce the frustration of manual post-editing.

Versatility and Potential Trade-offs

Gemini 3.5 Transcribe arrives with robust language support, functioning in 85 different languages. It also includes capabilities for processing pre-recorded audio, with the ability to distinguish between up to three different speakers.

However, the shift toward “intelligent” transcription introduces a new set of considerations. By design, the AI is tasked with interpreting and polishing speech, which means it technically alters the literal wording of what the user says. While this makes for cleaner, more readable notes and messages, it may not be suitable for contexts where a verbatim record is strictly required. For everyday communication, however, the ability to turn a stream-of-consciousness thought into a polished paragraph is expected to become an essential tool for Google users looking to streamline their mobile productivity.