在说话人识别中，当存在两个或多个发声类似的说话人时，会导致错误识别。为了提高在这种情况下的识别准确率，在音素层次上找出说话人特有的特征，将这些特征的子集构成一个该说话人特有的特征集，然后在这些特征集的基础上用GMM和i-矢量的方法对说话人进行识别。在实验室环境下收集了50个说话人的声音，分别在不同信噪比的环境下进行测试。实验结果表明提出的方法能够提高当存在发声类似的说话人时的识别准确率。%In speaker recognition tasks, one of the reasons for reduced accuracy is due to closely resembling speakers in the acoustic space. In order to improve the recognition accuracy in this condition, this paper proposes a technique by finding speaker-specific features at the level of phonemes and formulate a text using the subset of features that are unique, for speaker recognition task using GMM-based approach and i-vector based approach. Collecting 50 speakers’ speech in the labtorary environment and test them under different Signal-to-Noise Ratios. The experiments show that the proposed method can improve the recognition accuracy when there are closely resembling speakers.
展开▼
