Target Speaker Tagging benchmark tackles multi-conversation speaker ID

A team of researchers (Minjae Lee, Hee-Soo Heo, et al.) has proposed Target Speaker Tagging (TST), a new task that merges speaker diarization, verification, and identification into a single workflow. Given long audio recordings and a set of pre-enrolled speakers, a TST system must detect and label speech segments for known speakers while rejecting unknown voices. This problem is highly practical for meeting transcription, security, and multi-party analytics, yet until now lacked a standardized evaluation resource.

To fill that gap, the authors introduce TST-Bench, a large-scale synthetic benchmark with over 150 enrolled speakers, 300 conversation sessions ranging from 20 to 60 minutes, and reference annotations using global speaker labels. They define an evaluation protocol covering both pure diarization and full-pipeline scenarios. Experiments on real and synthetic data confirm that TST poses unique challenges beyond conventional benchmarks, and that purpose-built architectures significantly outperform naive combinations of existing diarization and verification systems. The dataset and protocols are publicly released.