CAJ | 학술논문

为了进一步提高基于传统的GMM模型的说话人辨识的识别率，引入了GMM?UBM模型，并且在特征提取方面采用多种特征参数组合来代替单一特征参数，以提高有效特征维数来弥补特征样本的不足，同时在说话人辨识的端点检测部分，用基于MFCC相似度和谱熵的端点检测方法来代替传统的基于短时能量和过零点的方法，以解决其对含噪语音检测不准确而影响说话人辨识的问题。实验表明，与传统的GMM模型相比， GMM?UBM模型能够有效地提高说话人辨识的性能，并且使用组合特征参数和利用基于MFCC相似度和谱熵的端点检测方法都可以进一步提高说话人辨识的性能。
위료진일보제고기우전통적GMM모형적설화인변식적식별솔，인입료GMM?UBM모형，병차재특정제취방면채용다충특정삼수조합래대체단일특정삼수，이제고유효특정유수래미보특정양본적불족，동시재설화인변식적단점검측부분，용기우MFCC상사도화보적적단점검측방법래대체전통적기우단시능량화과영점적방법，이해결기대함조어음검측불준학이영향설화인변식적문제。실험표명，여전통적GMM모형상비， GMM?UBM모형능구유효지제고설화인변식적성능，병차사용조합특정삼수화이용기우MFCC상사도화보적적단점검측방법도가이진일보제고설화인변식적성능。
To further improve the performance of speaker identification based on GMM model,GMM?UBM model is introduced into this paper.The feature combination instead of single feature is used to increase effective dimensions of speech features.The voice ac?tivity detection based on MFCC similarity and spectrum entropy instead of short?time energy and zero?crossing counts is used to effec?tively detect the voiced speech.Experimental results show that the performance of speaker identification based on GMM?UBM model is better than that based on GMM model.Both Feature combination and the voice activity based on MFCC similarity and spectrum entropy can further improve the performance of speaker identification.