Hang Chen, Shilong Wu, Jia Pan, Sabato Marco Siniscalchi, Jun Du · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3838658
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Speaker diarization technology is still challenged by poor cross-domain generalization of deep learning models, largely due to the limited availability of large-scale data resources. To address this issue, we present MISP-M 3 SD, a large-scale multimodal, multi-scenario, and multilingual dataset for robust speaker diarization. Specifically, MISP-M 3 SD contains more than 770 hours of synchronized audio-visual recordings collected from in-the-wild online videos, covering 14 scenarios and 16 languages. To enable efficient, cost-effective, and scalable dataset construction, we develop a largely automated pipeline including multilingual multi-scenario media acquisition, data cleaning and preprocessing, and cross-modal consistency-guided annotation. Specifically, we compare the predictions of audio-only, visual-only, and audio-visual diarization branches, and route low-confidence cross-modal inconsistency cases to selective manual verification. Experiments in both audio-only and audio-visual settings show that MISP-M 3 SD is an effective training resource for cross-dataset generalization, and also provides a strong initialization for efficient scenario-specific adaptation with limited in-domain data. All related resources are publicly available at https://github.com/coalboss/MISP-M3SD.
No comments yet — start the discussion below.