一种基于模式识别的多路盲语音提取方法

盲分离算法能在缺少混合系统参数的条件下仅由观测信号估计初始源,但分离信号存在固有的排列模糊性,这往往导致两次批处理过程中同一信号"对不准",因此很难获得连续的源信号。本文针对盲声源分离中存在的相同问题,根据语音和其他音频信号的特征差异,提出一种修正的自相关函数并以其值作为一个特征基元来表征声音信号的时序相关特性,同时用平均声门波形状参数作为另一个特征基元来表征语音产生的生理效应。以这两个参数作为识别不同音频信号的二维模式特征,采用一种模糊聚类算法提取多路盲分离语音。本方法有效克服了批处理盲声源分离中的信号排列顺序的不确定性,并通过选择合适的阈值提取多路连续语音。仿真给出了5路混合音频信号中盲提取两路连续语音的实验结果。

An approach to multiple blind-speech-abstraction based on pattern recognition

Classic blind source separation algorithm can estimate the original sources without given parameters of the mixture system by only using observation signals,but the inherent permutation ambiguity always leads to a signal not being"end to end"during the two batch processings,and so it is difficult to obtain the consecutive sources.Aiming at the same question in the blind audio source separation,this paper proposes a pattern rec- ognition approach for its solution.From the difference between the characteristics of speech and other audio signals,a modified autocorrelation function is proposed and its val- ue is taken as a characteristic basis to express the time sequence correlation of different au- dio signals,while the mean glottal flow shape parameter is taken as the other characteristic basis to express the physiological effect that speech produces.The above parameters are selected as two-dimension pattern characteristic to distinguish different audio signals,and a fuzzy clustering algorithm is used to extract multiple speeches blindly separated.This approach effectively solves the permutation ambiguity in batch blind audio source separa- tion and by selecting appropriate threshold can extract multiple consecutive speeches.Sim- ulation shows the experiment results of blind extracting two consecutive speeches among five mixture acoustic signals.