易扩展小样本环境说话人辨认系统的研究

Research on Expansile Speaker Identification System Adapted to Small-Scale Samples

针对基于人工神经网络的说话人辨认系统所存在的问题,提出了一种基于模糊最小二乘支持向量机(LS-SVMs)技术的两级分类说话人辨认系统。第一级对说话人进行大类预分,第二级则在大类范围内实现具体说话人的辨认。当有新的说话人加入时,只需要增加与新说话人相关的若干个二级分类器。系统仿真实验表明,在训练样本时长取3-11s,说话人由32人增至36人时,本文方法实现的系统训练时间明显降低,识别率更高。

Owing to the drawbacks of ANN-based speaker identification system, a novel speaker identification approach was proposed based on fuzzy LS-SVMs by hierarchical classify. In the first level, speech was classified to the rough class, and during the second level, the actual identification could be got. When speaker number increased, nothing but adding some corresponding classifiers. The experiments show that, when the training speech is 3-1 Is, the number of speakers change from 32 to 36, and the proposed system can not only reduce the training time, but also improve the recognition rate.