Multi-branch Learning for Noisy and Reverberant Monaural Speech Separation

Abstract

With the rapid development of deep learning approaches, much progress has been made on speech enhancement, speech dereverberation, and monaural multi- speaker speech separation to solve the cocktail problem. Some excellent methods have been proposed to solve the monaural speech separation in a noisy and reverberant environment. However, few studies exploit the correlations between anechoic speech and reverberant speech. In this work, the structure of a popular separation system is deconstructed, and a multi-branch learning method is proposed to enforce the network to exploit the correlations between anechoic speech and the corresponding reverberant speech. The results show that using multi-branch learning can improve the separation performance of different networks by 0.7dB with conv-tasnet on the WHAMR! dataset.