Speaker identification using multipulse excited linear prediction analysis (LPC) vocal tract model
暂无分享,去创建一个
The authors devised a method to identify a speaker where feature parameters are derived by normalizing the position and the magnitude of the subpulses in the multipulse train extracted from the vowel and are combined with 14th-order linear prediction analysis (LPC) cepstrum. In other words, 10 code vectors are constructed for each registered speaker in the 16-dimensional space. the speaker is identified by determining the code with the smallest VQ distortion for the vector of the unknown speaker.
A speaker identification experiment is executed using five sustained vowels of 100 speakers for the same time period. A 100-percent identification rate is realized by combining the feature parameters of the multipulse source composing the feature vector and the feature parameters of the LPC cepstrum by the weights of 3:2. the speaker identification also is attempted for 10 subjects after 10 days by the same method, and the identification rate of 69 percent is obtained. the result is an improvement of 12 percent, compared to the case where only the LPC cepstrum is used. the identification rate also is improved for the speech one and two months later.
[1] B. Atal. Automatic Speaker Recognition Based on Pitch Contours , 1969 .
[2] Bishnu S. Atal,et al. A new model of LPC excitation for producing natural-sounding speech at low bit rates , 1982, ICASSP.
[3] Robert M. Gray,et al. An Algorithm for Vector Quantizer Design , 1980, IEEE Trans. Commun..