Finetune Large Pre-Trained Model Based on Frequency-Wise Multi-Query Attention Pooling for Anomalous Sound Detection

Existing anomalous sound detection (ASD) methods generally rely on the learned patterns of normal sounds to identify deviations as anomalies in generative and discriminative ways. However, partly due to the diverse and sparse distribution of anomalies, they may suffer from poor generalization capability to distinguish unseen normal sounds and anomalies. In this paper, we propose to finetune the large pre-trained model based on frequency-wise multi-query attention pooling (FMQAP) for ASD. Specifically, FMQAP first dynamically fuses multiple output embeddings from the pre-trained model, and then exploits a novel frequency-wise attention pooling scheme to improve the generalization capability for various machine types. Evaluations on DCASE 2023 Task2 benchmark demonstrate the superiority of FMQAP, achieving the state-of-the-art 77.62% in terms of official score.