音源特征用于提高话者确认系统的鲁棒性

Integrating sound source features for robust speaker verification

给出了一种采用音源特征信息提高以声道倒谱参数为特征的话者确认系统噪声鲁棒性的方法,提取了两类音源特征参数:短时(单帧)特征参数和较长时(多帧)特征参数,并分别构建了两个利用音源特征参数的与文本无关的话者确认辅助子系统.采用线性加权对主、辅子系统的输出进行融合.在NIST’03数据库上100个男性话者的对比实验表明,音源特征参数具有良好的噪声鲁棒性,声道特征与音源特征具有较强的互补性,尤其是在较强的噪声背景下,利用音源特征可以有效地提高以声道倒谱参数为特征的确认系统的鲁棒性.

Two approaches that use the Pitch and Energy of sound source to improve the performance of speaker verification were presented. The first approach extracted speaker-dependent features of the sound source by short-term analysis. The second approach captured the Pitch and Energy dynamics by taking into account the statistics of multi-frames. For each feature, a model was established by means of GMM-UBM, and then the weighted linear sum of scores from each model was computed as the fusion score. Experiments conducted on the NIST＇03 database show that sound source features have a complementary effect on MFCC. Further more, these features are more robust than MFCC and can greatly improve robustness in noisy environments.