The "answer only when spoken to" era of audio AI is over. Meet a model that keeps listening to sound, environment, and instructions — and acts on its own 🎧 Title: Audio Interaction Model URL: 🎧 Overview A unified, always-on streaming Large Audio Language Model (LALM). It runs a continuous perceive-decide-respond loop, listening to sound, environmental audio, and user instructions at once, then responding dynamically based on semantic understanding of the stream. ❓ Challenges Solved Most current LALMs run offline and handle isolated tasks (streaming ASR or voice chat separately). Real interaction needs an always-on model that listens in real time and reacts on the fly. 💡 Methodology & Proposed Approach The core is the SoundFlow framework operationalizing the perceive-decide-respond loop. ・Streaming-native data construction ・Comprehension-aware training ・Asynchronous low-latency inference for stable real-time interaction It trains on StreamAudio-2M (2.6M items) covering 7 fundamental abilities and 28 sub-tasks, plus Proactive-Sound-Bench to assess proactive intervention. 📊 Experimental Results / Use Cases ・Maintains competitive performance across 8 benchmarks ・Enables capabilities offline LALMs can't: real-time ASR, streaming audio instruction following, and proactive intervention Great for always-on voice assistants, real-time dialogue, and proactive audio assistance. #AudioAI# #LALM#