Tsinghua Open-Sources Falcon, a Lightweight Lip-Reading Speech Extraction Model

Tsinghua's team open-sourced Falcon, a NeurIPS 2026 audio-visual target speaker extraction model that extracts a specific person's voice in real time from noisy multi-speaker environments using lip movements. With only 0.5M parameters, it cuts parameters by 91.9% and latency by 11x, with code, models, a demo and the RealSSA benchmark released.

2026-10-31 ~ 2026-10-31 · 2 related posts

• Tsinghua Open-Sources Falcon: Visual Target Speaker Extraction With 11x Lower Latency — aigclink · 2026-10-31
• Tsinghua's Falcon Hits NeurIPS: 0.5M-Param Streaming Lip-Guided Voice Extraction + RealSSA Benchmark — aigclink · 2026-10-31