0.4秒ごとに聴き、話すオープンソースモデル
研究者らはAudio-Interactionをリリースした。0.4秒ごとのチャンクで音声を処理して発話のタイミングを判断する3Bのオープンソースモデルで、ノイズ検出ではGemini 3 Flashを上回る性能を示した。
researchers released audio-interaction, an open-source 3b model designed to unify speech recognition, translation, and dialog by processing audio streams in real-time. unlike turn-based models like gpt-4o, it segments audio into 0.4-second chunks and outputs a <silent> or <response> token after each, enabling natural turn-taking and simultaneous listening. trained on the synthetic streamaudio-2m dataset of 302,000 hours, the system beats gemini 3 flash in proactive noise detection and matches 7b models on the mmau benchmark. this approach addresses the limitations of current streaming models that cannot handle multiple tasks or react proactively to environmental sounds, signaling a shift toward more capable, multimodal voice agents.