Single-channel Speech Enhancement Student under Multi-channel Speech Enhancement Teacher

Abstract

In recent years, significant success has been made in single-channel speech enhancement using the deep neural networks. These approaches trained a model on synthetic noisy speech corpus, which was created by adding noise to clean speech. Because there is a mismatch between synthetic training data and the actual application environment, the model's performance is not guaranteed. This paper proposes to use a multi-channel speech enhancement teacher model to guide a single-channel noise suppression student model. We set the multi-channel teacher's processed signal as the single-channel student's training target. With our proposed approach, the single-channel speech enhancement model can be trained using real noisy speech and performed as well as a multi-channel speech enhancement model. Experimental results on CHIME-3 demonstrate that our proposed approach can achieve competitive performance both in speech enhancement and automatic speech recognition tasks, even without ground truth signals.