Acoustic and auxiliary speech features for speaker identification system

摘要

The focus of the article is on the selection, adjustment and overall performance of speech features at acoustical and prosodic level for speaker recognition task. Namely: perceptual linear prediction, Mel frequency cepstra, cepstral linear prediction, formant frequencies, and different auxiliary features. Both brief theoretical backgrounds and possible computational methods are outlined in regard to the speaker recognition task. In the series of experiments using 114 speakers database, it was observed that a model based method slightly outperformed the perceptual ones. Furthermore, it was found that auxiliary and prosodic features may not always improve scores when processed together with acoustic ones. On average the success rate was about 90% whereas the best recorded score was 99.1% for cepstral linear prediction coefficients in connection with k-nearest neighbor classifier.