Missing data recovery using autoencoder for multi-channel acoustic scene classification

In this paper, we propose a method of missing data recovery using an autoencoder for multi-channel signals. Recently, many deep neural network-based classification methods using multi-channel signals have been proposed. The advantage of using multi-channel signals is that both frequency and spatial information can be used. However, systems using such signals are vulnerable to missing data because of a mismatch between the training and testing data. To minimize the mismatch, some techniques that include simulated missing data to the training data have been proposed. However, it is difficult to prepare missing data covering all possible mismatch situations. Therefore, we focus on using an autoencoder to recover the missing data without any assumptions of missing situations. In the case of multi-channel data inputted into an autoencoder, channel relationships are compressed into a low-dimensional hidden layer. Then, the autoencoder outputs data to reconstruct the input data from the layer. Therefore, when multi-channel input into the autoencoder has some missing channels, the output of the autoencoder is expected to recover some missing channel information by using the hidden layer. Since the autoencoder is regarded as a preprocessing of acoustic tasks, we evaluated the proposed method using an acoustic classification task. From the experimental results, we confirmed that the proposed method can recover missing data and improve the classification performance.