Investigation on Spatial and Frequency-Based Features for Asynchronous Acoustic Scene Analysis

We investigate the characteristics of spatial and frequency-based features as an aspect of asynchronous acoustic-scene analysis. Recently, many tasks related to acoustic scene classification (ASC) have attracted a lot of interest. One of the most famous events, the challenges and workshops on detection and classification of acoustic scenes and events (DCASE) series, started in 2013. Here, we focus on DCASE 2018 Task 5, a multi-channel dataset, from among the many tasks of the challenges. This dataset consists of audio data acquired by multiple mi-crophone arrays at different positions. It should provide rich spatial information on acoustic scenes. However, little research has exploited spatial features for the task. We focus on the spatial cepstrum, a robust spatial feature. First, we analyze DCASE 2018 Task 5 with cross-correlation to estimate synchronous and asynchronous data. Then, a convolutional neural network-based ASC model using the spatial cepstrum is used to evaluate performances under synchronous or asynchronous datasets. In our experiments, a system with spatial cepstrum and frequency-based features was evaluated under multi-channel datasets, and the results show that it achieved the highest F-score.