Voice source cepstrum processing for speaker identification

Voice source analysis and modelling has played a key role in important speech applications such as speech recognition, speech synthesis and speaker recognition. This work presents a robust algorithm for glottal closure detection and a novel set of voice source features for speaker recognition. In the rst part of the dissertation the DYPSA algorithm is developed for detecting glottal closure instants (GCIs). It includes a detailed study of group delay functions and their application to the linear prediction residual; glottal closure candidate generation from the group delay function; cost function design with regards to the properties of the speech signal at the point of closure; and dynamic programming algorithm used to reject unlikely glottal closure candidates. The DYPSA algorithm is evaluated on a speech database that includes simultaneous laryngograph recording to provide reference glottal closures instants. The algorithm achieves a 95.7% identi cation rate with 0.71 ms timing error standard deviation. In the second part of the dissertation GCI detection allows the vocal tract transfer function to be estimated using closed-phase analysis. This is converted to cepstrum coe cients (VTCC) and subtracted from the mel-frequency cepstrum coe cients (MFCC) to derive a set of voice source cepstrum coe cients (VSCC). These are then used for speaker identi cation on the TIMIT database. We show that although a classi er using MFCC performs better than one using VSCC, the combination of the two gives a signi cant improvement in recognition rate, illustrat-