一种新的基于稀疏分解的单通道混合语音分离方法

A New Single-Channel Speech Separation Method Based on Sparse Decomposition

论文以新的语音信号稀疏基—准KLT基的构造为基础,提出了一种新的基于稀疏分解的单通道混合语音分离方法.论文首先以理想准KLT基的构造为基础,从理论上提出并证明了基于各源语音信号的理想准KLT基,利用lexp(0)-范数优化算法,可实现单通道混合语音的完美分离.鉴于单通道混合语音分离时,无法精确求取各源语音信号的理想准KLT基,论文提出先基于正交匹配追踪算法,以混合语音信号为已知条件,构造各源语音信号的正交匹配追踪模板匹配准KLT基,再由lexp(0)-范数优化算法来分离单通道混合语音.仿真实验表明论文所提理论的正确性,和基于正交匹配追踪模板匹配准KLT基来分离单通道混合语音信号的有效性.

This paper proposes a new single-channel speech separation method based on sparse decomposition. The new method is based on the construction algorithms of quasi-KLT bases proposed in this paper.First,it is proved that based on the ideal quasi-KLT bases generated from the sources,both sources can be perfectly separated from a single mixture by l exp（0） optimization. Then, considering the fact that the ideal quasi-KLT bases cannot be obtained in practice, orthogonal matching pursuit （OMP） template-matching quasi-KLT bases generated from the single mixture are first proposed, and then perform single-channel speech separation by 1 exp（0） optimization using the OMP template-matching quasi-KLT bases. Finally, simulation results demonstrating the perfect separation of both sources from a single-channel speech mixture by 1 exp（0） optimization using ideal quasi-KLT bases, as well as the effective performance of single-channel separation of two speech sources using the OMP template-matching quasi-KLT bases are presented.