基于自组织聚类和改进粒子群算法的语音转换方法

Voice conversion based on self organization clustering and modified particle swarm optimization

提出一种基于自组织聚类,并且利用改进粒子群算法确定转换模型参数的语音转换方法.该方法首先基于自组织特征映射网络对特征参数进行聚类,再对每个聚类分别建立转换规则,并且利用柯西变异的粒子群算法确定每个转换规则中的模型参数.与传统的单一转换规则相比,聚类后建立的多转换规则以及利用改进粒子群算法确定参数能够提高映射关系的准确度,避免参数陷入局部最优点。以女声转男声为例,主观测试表明该方法得到的转换语音与目标的相似度提高了27.6%,平均主观意见分(Mean Opinion Score,MOS)提高了0.6,客观测试也表明该方法谱失真最小,与目标的包络更接近.

A method of voice conversion based on self organization clustering and determining parameters of conversion model by modified Particle Swarm Optimization （PSO） is proposed. Firstly, Self Organization Feature Mapping （SOFM） Network is used to cluster the characteristic parameters, and then the conversion rule for each cluster is established, where the parameters of conversion model in each conversion rule are determined by modified PSO with Cauchy mutation. Compared with the single conversion rule in conventional method, the multiple rules established by using clustering and parameters determined by modified PSO can improve the accuracy of mapping function and avoid the model parameters being trapped in the local optimum. The experiments take the conversion from female to male as example, by subjective evaluation, the proposed method increases the similarity by 27.6% through ABX test, and increases the Mean Opinion Score （MOS） by 0.6, and by objective evaluation, the spectral distortion with proposed method is the least.