CAJ | 학술논문

为了提高说话人识别系统的识别效率，提出一种基于说话人模型聚类的说话人识别方法，通过近似KL距离将相似的说话人模型聚类，为每类确定类中心和类代表，构成分级说话人识别模型。测试时先通过计算测试矢量与类中心或类代表之间的距离选择类，再通过计算测试矢量与选中类中的说话人模型之间对数似然度确定目标说话人，这样可以大大减少计算量。实验结果显示，在相同条件下，基于说话人模型聚类的说话人识别的识别速度要比传统的GMM的识别速度快4倍，但是识别正确率只降低了0.95%。因此，与传统GMM相比，基于说话人模型聚类的说话人识别能在保证识别正确率的同时大大提高识别速度。
위료제고설화인식별계통적식별효솔，제출일충기우설화인모형취류적설화인식별방법，통과근사KL거리장상사적설화인모형취류，위매류학정류중심화류대표，구성분급설화인식별모형。측시시선통과계산측시시량여류중심혹류대표지간적거리선택류，재통과계산측시시량여선중류중적설화인모형지간대수사연도학정목표설화인，저양가이대대감소계산량。실험결과현시，재상동조건하，기우설화인모형취류적설화인식별적식별속도요비전통적GMM적식별속도쾌4배，단시식별정학솔지강저료0.95%。인차，여전통GMM상비，기우설화인모형취류적설화인식별능재보증식별정학솔적동시대대제고식별속도。
This paper proposes a speaker recognition method based on Speaker Model Clustering(SMC)to improve the efficiency of the recognition system. Through the calculation of an approximated Kullback-Leibler divergence, the similar speaker model is clustered. All of cluster centroid and cluster representative construct a hierarchical speaker recognition model together. During the recognition stage, the cluster is selected by calculating distance between the test vectors and cluster centroids or cluster representatives on the first step. In accordance with calculating the logarithmic likelihood between the test vectors and the speaker models in the selected cluster, the speaker is determined, with the sharp decrease-ment of computation. The experimental results show that the proposed method improves the recognition speed about four times and loses the accuracy rate about 0.95%compared with the traditional Gaussian Mixture Model(GMM). In conclu-sion, the SMC method can improve the recognition speed with almost the same accuracy.