针对说话人分段与聚类算法中先验知识不足的问题，利用基于信息瓶颈(IB)准则和基于隐马尔科夫模型(HMM)/高斯混合模型(GMM)方法间的互补性，提出了一种基于特征层融合的说话人分段与聚类算法.该算法将基于 IB准则算法的输出结果进行对数变换和降维处理；然后利用变换后的特征与传统梅尔频率倒谱系数(MFCC)特征分别训练说话人 GMM模型，并在得分域对说话人类别的得分进行加权融合；根据融合的得分，进行基于 HMM/GMM 模型的说话人分段与聚类.实验表明，融合后的特征可以为系统提供更多的先验信息，比传统方法的误配率降低了1.2%.%The performance of the speaker segmentation and clustering system usually degrades because of lacking prior information about the speakers. To solve the problem, a novel approach that combines the algorithms based on Information Bottleneck(IB) principle and Hidden Markov Model(HMM)/Gaussian Mixture Model(GMM) is proposed by using the complementarity of these two algorithms. After logarithmic transform and Principal Component Analysis to reduce dimension, the output of the IB algorithm is then used to train the speaker GMM model. Along with the speaker GMM model trained by the traditional Mel Frequency Cepstral Coefficient(MFCC) feature, the scores between different speaker clusters are computed respectively and then combined using linear weighted sum method. Lastly, the HMM/GMM based speaker segmentation and clustering is performed with the combined score. Experiments show that the IB features provide more prior information for the system and the speaker match error rate is reduced by 1.2%compared to that in traditional method.
展开▼
