Unsupervised Anomalous Sound Detection Using Timbral and Human Voice Disorder-Related Acoustic Features

Anomalous sound detection (ASD) crucially prevents industrial accidents by distinguishing normal and abnormal machine sounds. Previous research utilizing timbral and short-term features attained a notable F1 score of 0.920. However, relying solely on supervised learning models is impractical due to the difficulty of acquiring anomaly data. This study focuses on developing an unsupervised learning model for ASD, emphasizing prominent timbral features. We also investigate human voice disorder (HVD)-related features, which are potentially linked to human perception of anomalous sounds in machines. We conducted a comparative analysis using 5-fold cross-validation to evaluate our proposed method, with the area under receiver operating characteristic (ROC AUC) as the metric. The proposed ASD method using timbral and HVD-related features significantly improved AUC by 10.87% compared to the baseline system in the DCASE Challenge 2020.