An Ensemble Method for Multiple Speech Enhancement Using Deep Learning

This paper proposes an ensemble of multiple speech enhancement methods using convolutional neural networks (CNN). Speech enhancement is one of the most important tasks in the field of audio signal processing, and various methods have been proposed so far. Each of these methods has its own strengths and weaknesses depending on the target speech signal, the type of noise, and the recording environment. In this study, we aim to construct a novel robust speech enhancement method that works well in various environments by combining the advantages of multiple speech enhancement methods. We formulate an ensemble based on weighted summation of time-frequency masks, and propose a method to estimate the optimal weight values based on the input acoustic signal using a convolutional neural network. The convolutional operation in CNN allows the integration to take into account the changes in time and frequency. In the simulation experiments, we evaluate the effectiveness of the proposed method by computing the mean square error between the ideal mask and the mask generated by each method.