Anomalous Sound Event Detection Based on WaveNet

This paper proposes a new method of anomalous sound event detection for use in public spaces. The proposed method utilizes WaveNet, a generative model based on a convolutional neural network, to model in the time domain the various acoustic patterns which occur in public spaces. When the model detects unknown acoustic patterns, they are identified as anomalous sound events. WaveNet has been used to precisely model a waveform signal and to directly generate it using random sampling in generation tasks, such as speech synthesis. On the other hand, our proposed method uses WaveNet as a predictor rather than a generator to detect waveform segments causing large prediction errors as unknown acoustic patterns. Because WaveNet is capable of modeling detailed temporal structures, such as phase information, of the waveform signals, the proposed method is expected to detect anomalous sound events more accurately than conventional methods based on reconstruction errors of acoustic features. To evaluate the performance of the proposed method, we conduct an experimental evaluation using a real-world dataset recorded in a subway station. We compare the proposed method with the conventional feature-based methods such as an auto-encoder and a long short-term memory network. Experimental results demonstrate that the proposed method outperforms the conventional methods and that the prediction errors of WaveNet can be effectively used as a good metric for unsupervised anomalous detection.