Solution of Underdetermined Speech Separation Problems by Combining ICA and Time-Frequency Masking Methods

摘要

Blind source separation of speech signals has many potential applications. Toward practical applications, underdetermined problem attracts much interest. This paper proposes a new approach to solve underdetermined problem. There are three stages in the approach: 1) Separate mixtures into two independent parts; 2) Two time-frequency masks are then generated; 3) Conventional time-frequency masking method is applied to separate each source. Experimental results prove the effectiveness of the proposal. Compared with conventional methods, the proposal improves signal-to-interference ratio by at least 4～5 dB in the same setup. The advantages of the proposed method are simple array configuration, low computation cost, and higher separation performance.