Spectral-to-Spatial Distillation: Denoising Framework for Real-Time Anomalous Sound Detection

In this paper, we propose spectral-to-spatial distillation, a novel denoising framework for real-time anomalous sound detection. While anomalous sound detection is crucial for industrial applications, its reliability is often compromised by background noise, which can lead to false positives. Our proposed method addresses the issue of background noise by distilling knowledge from a general-purpose spectral filtering network into an environment-specific spatial filtering network. Specifically, we generate distillation targets, which are audio signals with reduced noise, using a pre-trained foundation model. A spatial filtering network is then trained using these targets. A key feature of our distillation process is its ability to automatically generate these targets using only one-shot, brief, noise-free reference signal of the target sound. Furthermore, we introduce a new quality metric for these distillation targets, called Semantic Clarity improvement (SCi). By leveraging the semantic audio embedding capabilities of a foundation model, SCi measures the improvement in semantic similarity between the distillation target and the reference signal. This SCi allows for effective distillation by weighing the loss function based on the quality of the targets. Experimental results demonstrate that our method achieves the best denoising and anomaly detection performance while maintaining real-time processing capabilities, making it a practical solution for noisy industrial environments.