P6-07: Self-Refining of Pseudo Labels for Music Source Separation With Noisy Labeled Data
Junghyun Koo (Seoul National University), Yunkee Chae (Seoul National University)*, Chang-Bin Jeon (Seoul National University), Kyogu Lee (Seoul National University)
Subjects (starting with primary): MIR tasks -> automatic classification ; MIR tasks -> sound source separation ; Evaluation, datasets, and reproducibility -> annotation protocols
Presented In Person: 4-minute short-format presentation
Music source separation (MSS) faces challenges due to limited availability and potential noise in correctly labeled individual instrument tracks. In this paper, we propose an automated approach for refining mislabeled instrument tracks in a partially noisy-labeled dataset. The proposed self-refining technique with noisy-labeled dataset results in only a 1% accuracy degradation for multi-label instrument recognition compared to a classifier trained with a clean-labeled dataset. The study demonstrates the importance of refining noisy-labeled data for training MSS models and shows that utilizing the refined dataset for MSS leads to comparable results to a clean-labeled dataset. Notably, upon only access to a noisy dataset, MSS models trained on self-refined datasets even outperformed those trained on datasets refined with a classifier trained on clean labels.
If the video does not load properly please use the direct link to video