잡음환경에 강인한 화자확인에 관한 연구

Robust Speaker Verification in Noisy Environments

The first proposed speaker verification method for the preprocessing part is the linear predictive coding (LPC) cepstrum weighting function that modifies the existing one. Simulation results show that the average speaker verification rate is improved by 17.7% using the proposed end-point detection algorithm using LPC residue, and it is improved by 37% using the proposed noise cancelling and microphone property compensation algorithm. The proposed weighting function for discriminating inter-speaker variations also improves the average speaker verification rate by 6.6% (FR). The second proposed speaker verification method for the speaker verification part is a modified clustering algorithm. In this study, a nonlinear clustering algorithm is used. The proposed method increased the speaker verification rate by 1.0% (FR). The following are proposed in the study: a method for reducing the training utterance number, a method using the personal vector quantization code book for improving the imposter acceptance rate, the observation symbol probability smoothing method by probability weights for improving the DHMM performance, and the speaker model adaptation method for adapting the voice variance. The two verification methods mentioned above are applied to the general DHMM. The results showed a performance enhancement of about 1.1% in FA and 41.7% in FR.