Improving Anomalous Sound Detection with Top-M Pseudo-labeling
Recent years have witnessed the growing popularity of discriminative self-supervised learning methods for learning effective machine sound embeddings in the Anomalous Sound Detection (ASD) task. These methods typically incorporate auxiliary tasks such as machine attribute classification to guide representation learning. In practice, when ground-truth metadata is limited, pseudo-labeling has emerged as an approach to supply supervisory labels for such auxiliary tasks. Existing pseudo-labeling strategies for ASD include approaches that assign a single most confident attribute to each instance. Complementary to these methods, this work explores a Top-M pseudo-labeling strategy to enhance discriminative self-supervised learning for ASD tasks. Instead of relying solely on the only attribute prediction, we incorporated the Top-M most confident attributes after pseudo-label generation to better capture semantic uncertainty. We proposed a categorization-based framework that adapts pseudo-labeling to attribute availability. Experiments were conducted on ASD benchmark datasets (DCASE 2020–2025) and the results on DCASE 2025 development test dataset showed that Top-M labeling consistently outperformed Top-1 across models. These results highlight the general effectiveness of our approach in improving ASD performance under limited supervision.
