TL;DR: Falcon, a 0.5M-parameter model, does streaming audio-visual target speaker extraction using lip video, with about 20x less compute than AV-TF-GridNet. The paper also releases RealSSA, a more realistic evaluation set. Efficient Streaming Audio-Visual Target Speaker Extraction for Real-World Acoustic Scenes Key points ・🎧 Existing benchmarks mostly use two speakers and anechoic mixes, missing reverberation and near/far-field interference ・🏠 RealSSA-Sim covers four room types with 1–5 near-field and 10–15 far-field interferers (about 153 hours), plus a real phone-recorded test set ・🧩 Falcon is a single-pass U-Net separator that models time-frequency only at the bottleneck and injects global context into skip connections via GCI ・🎯 A training-only speaker contrastive loss pushes outputs away from near-field interferers, with no added inference cost ・⚡ 10.6G MACs and 51.3 ms GPU latency, versus 45.3G and 172.1 ms for SwiftNet-12 ・📈 Best SI-SNRi and SDRi on all four simulated benchmarks, though AV-Mossformer2 edges it on real-recording WER It balances efficiency and accuracy while making evaluation more realistic. The loss needs clean interferer waveforms, so applying it to real recordings is future work. #SpeechProcessing# #AudioVisual#