两步单通道语音分离算法

Two-step single channel speech separation algorithm

针对单通道语音分离算法未充分利用语音的相位信息,导致分离的性能不佳,提出了两步单通道语音分离算法。利用特征提取网络作为编码器,波形重构网络作为解码器进行预训练;加入分离模块,利用该模块对提取的声学特征进行操作,估计独立语音信号的掩模。此外,基于TIMIT语音数据集进行仿真实验,与排列不变训练(Permutation invariant training,PIT)算法及其改进的算法进行对比。实验结果表明,提出的方法在信噪失真比(Signal to noise distortion ratio,SDR),信噪伪影比(Signal to noise artifact ratio,SAR),信噪干扰比(Signal to noise interference ratio,SIR)的结果更高,分离性能更优。

A two-step single channel speech separation algorithm is proposed because the current single channel speech separation algorithm does not make full use of the phase information of speech,resulting in poor separation performance.In the first step,the feature extraction network is used as the encoder and the waveform reconstruction network is used as the decoder for pre training.In the second,an additional separation module is used to operate on the extracted acoustic features to estimate the mask of independent speech signals.In addition,simulation experiments are carried out based on TIMIT speech data set to explore the performance of the proposed algorithm,and compared with the permutation invariant training(PIT)algorithm and its improved algorithm.The experimental results show that the proposed method has higher signal to noise distortion ratio,signal to noise artifact ratio and signal to noise interference ratio,and better separation performance.