Wendi Sang, Kai Li +3School of Computer Technology and Application, Qinghai University, Xining 810016, China, Department of Computer Science and Technology, Institute for AI, BNRist, Tsinghua Laboratory of Brain and Intelligence (THBI), IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing 100084, China +2
Extracts a target speaker’s speech from mixed audio using that speaker’s mouth video, with causal streaming inference. It introduces a single-pass multi-scale separator and the RealSSA benchmark for dynamic, realistic acoustic scenes.
AI Dataset2026
Preserves a model conversation trace and raw token IDs for investigating unexpected link generation by a Qwen model. Includes a readable trace export and Parquet data to support reproduction.
AI Dataset2026
Evaluates document parsers on 100–300-page documents, including cross-page text, tables, and layout structure. Its scale exposes failures that page-level benchmarks can miss.