Skip to main content

First-Shot Unsupervised Anomalous Sound Detection With Metadata-Assisted Audio Generation

By
Feiyang Xiao; Jian Guan; Kejia Zhang; Qiaoxi Zhu; Wenbo Wang; Haohe Liu; Wenwu Wang

First-shot unsupervised anomalous sound detection (ASD) aims to detect anomalies in new machine types using limited normal sounds. Its applicability is hindered by the lack of real anomalies for parameter tuning and the scarcity of target-domain normal sounds, both of which limit the robustness of detection models to domain shifts caused by variations in operating and environmental conditions. To address these challenges, we propose Gen-ASD, a metadata-assisted audio generation and detection framework that synthesizes anomalous sounds for parameter tuning during model validation and normal sounds for improved representation of target domain conditions. Machine metadata describe machine status, operating conditions, and environmental factors that shape acoustic characteristics. Leveraging these relationships, Gen-ASD employs a consistency-enhanced text-to-audio generator to learn metadata-acoustic relationships from reference machine types and synthesize normal and anomalous sounds for unseen machine types. Generated anomalies serve as surrogate validation data for tuning-dependent ASD methods, while generated normal sounds complement under-represented target-domain conditions to improve the robustness of detection models to domain shifts. These complementary capabilities enable detector training and anomaly-aware validation without real anomalies of new machine types. Extensive experiments demonstrate that Gen-ASD integrates seamlessly with both non-deep-learning and deep-learning ASD models in a plug-and-play manner. Further analyses validate the effectiveness of generated validation data for parameter tuning and the consistency of synthesised sounds with the under-represented target-domain conditions.

Read on IEEE Xplore