Falcon：高效流式音视频目标说话人提取方法
原标题：Paper page - Efficient Streaming Audio-Visual Target Speaker Extraction for Real-World Acoustic Scenes
AI 摘要
研究者提出轻量级因果音视频目标说话人提取方法 Falcon，通过单次编解码的多尺度时频分离与编码器空间说话人对比损失，在零推理成本下抑制近场干扰泄漏。同时发布 RealSSA 基准，包含可控模拟场景（1-5 个动态近场干扰说话人，远场场景另含 10-15 个远场干扰源）与真实录音留出评测集。Falcon 在 RealSSA-Sim、LRS2-2Mix、LRS3-2Mix、VoxCeleb2-2Mix 上取得最优提取质量，相比因果 AV-TFGridNet 基线参数量减少 91.9%、计算量减少 20.2 倍、GPU 推理延迟降低 11.1 倍。
正文节选
Efficient Streaming Audio-Visual Target Speaker Extraction for Real-World Acoustic Scenes Abstract Falcon is a lightweight causal audio-visual target speaker extraction method that achieves high quality with drastically reduced parameters and latency, supported by the realistic RealSSA benchmark. Streaming audio-visual target speaker extraction (AVTSE) is essential for latency-sensitive applications. However, current causal systems face two coupled gaps. First, standard benchmarks fix the number