Research on Unsupervised Campus Audio Anomaly Detection Based on Generative Adversarial Network

In order to detect inappropriate speech in classrooms, this paper proposes an unsupervised model based on generative adversarial networks (GAN), which integrates an autoencoder structure to detect anomalies in uncalibrated classroom audio, saving a lot of labeling time. Enhance data by adding Latent Space Pitting and weight sharing. Meanwhile, due to the previous model being mostly used for two-dimensional image data, in order to enhance the learning ability of one-dimensional speech information features, a multi head attention mechanism was introduced. Experiments were conducted on a self-made classroom audio dataset, demonstrating that the proposed method can improve the ability of anomaly detection and effectively distinguish anomalies.