UniCue: A Unified Benchmark for Multi-Cue Target Speaker Extraction with a Shared Conditioning Interface

Target speaker extraction (TSE) can be guided by heterogeneous cues such as video, enrollment speech, text, spatial information, etc. However, current multi-cue evaluation remains fragmented across cue settings, preprocessing pipelines, and fusion interfaces. To enable fair and systematic comparison, we establish UniCue, a unified benchmark for multi-cue TSE. UniCue provides a speech-in-video mixture dataset with paired cues from four modalities and speaker-disjoint splits, standardizes cue construction and preprocessing, and introduces modality-specific stress settings that explicitly probe cue availability and cue reliability. To reduce interface bias, we instantiate UniCueNet, a shared reference conditioning interface on a SepFormer-style separator backbone. The interface projects heterogeneous cues into a common latent space, organizes them into a unified cue memory, and injects the separator through mixture-guided attention over the cue memory, and injects cue information via layer-wise gated injection. We additionally validate the interface on a second separator family and observe that, although absolute performance varies across carriers, the main subset-wise trends remain consistent. Experiments over all cue subsets and selected stress conditions reveal clear patterns of cue usefulness, cross-cue complementarity, and robustness, establishing a reproducible testbed for multi-cue TSE research. Code and data will be released. Demo: https://uni-cue.github.io/UniCue-demo/