Estimation of AR model of vocal folds for speaker identification

Abstract

In this thesis, speaker identification is done by the AR model parameters of the vocal folds. Speaker identification needs extraction of speaker discriminative features. Melfrequency cepstral coefficients (MFCC) and Linear predictive cepstral coefficients (LPCC) are well known cepstral techniques which extract speaker discriminative vocal tract properties from speech signal for speaker identification purpose. On the other hand, the vocal folds properties of a speaker can also be used for this purpose as vocal folds vary person to person. But in this case the correct modeling of vocal folds is essential. AR model parameters of the vocal folds is used here as the speaker distinctive features. Vocal folds properties are found by inverse filtering the cepstral of the output speech by the vocal tract properties related LPCC. These model parameters called speaker features are then used to generate the so-called codebook of a speaker by the well established vector quantization technique. Codebooks generated in this way are then used to find the speaker identity using feature matching technique. The result of the proposed model found here is significantly better than that of the previous model for voiced sound.