一种基于衰减-时延聚类估计的欠定语音盲分离算法

Research on underdetermined speech blind separation based on attenuation and time-delay clustering estimation

针对衰减-时延混合模型,提出一种改进的欠定语音盲分离算法.第一步根据独立语音源在时频域上的稀疏性,用势函数法分别聚类估计衰减矩阵和时延矩阵,然后配对以确定各声源混合矩阵.第二步由估计的混合矩阵,采用改进最短路径法恢复出目标语音.为了减少计算量,设置门限对能量较小的时频点直接置零处理,在衰减矩阵和时延矩阵聚类估计时采用了分段聚类算法.仿真实验表明本文算法分离出的语音和源语音相似系数达0.96,0.97,0.93,信噪比达12.66dB,12.86dB,8.87dB,且有效减少了计算量,证明了该算法的可行性和有效性.

An improved algorithm of underdetermined speech blind separation is proposed for the attenu ation time-delay mixed model. The first step, according to the sparseness of independent voice sources in the time-frequency domain, attenuation matrix and time-delay matrix were Clustering estimated sepa- rately by using potential function method. Each voice source mixed matrix were determined by matching the two matrices. The second step, the target voice through the shortest path method was recovered ba- sing on the estimated mixed matrix. In order to reduce the amount of calculation, time-frequency points with lesser energy which can be regarded as noise were set zero by setting threshold and segment cluste- ring algorithm was proposed when clustered the attenuation matrix and time-delay matrix. Simulation results showed that similarity coefficient between the separated voice and source voice reached 0.96, 0. 97,0. 93 and SNR gain were 12.66dB, 12.86dB, 8.87dB by using the algorithm in this paper. The amount of calculation was reduced to some extent as well. Feasibility and effectiveness of the algorithm were proved.