Google Details Gemini 3.5 Audio Models and Native Audio Intelligence

Google has introduced "native audio intelligence," a technology that processes audio directly rather than relying on a multi-step text conversion pipeline. [1] This approach allows models like Gemini to understand both the sound and the intent simultaneously. [1] Gemini 3.5 Live Translate utilizes this to offer real-time speech translation across 70 languages and over 2,000 language combinations. [1] The system is designed to reflect emotional nuances and natural language transitions like code-switching. [1]

Google's Gemini 3.5 Transcribe operates in environments with heavy background noise and complex terminology to produce formatted text. [1] This model currently powers the "Rambler" feature in Android's Gboard, which removes filler words, corrects grammar, and enables voice-based text editing. [1] As part of its goal to support the world's 1,000 most used languages, Google trained its Universal Speech Model (USM) on 12 million hours of audio data. [1]

By moving away from traditional speech-to-text-to-speech pipelines, Google is treating audio as a primary modality, allowing models to capture non-verbal cues like tone and emotion. [1] The integration of these direct-processing models into everyday tools like Gboard and real-time translation services demonstrates a focus on practical, multimodal applications. [1] What the sources don't address: Whether the Gemini 3.5 audio features are currently available globally or rolling out in specific regions first.