ABSTRACT
Performance improvement of the speaker recognizers using the traditional methods like signal processing, has hit a dead end. Speech researchers are therefore now focusing on other techniques and processes to supplement the traditional methods to reduce the gap in communication interfaces between humans and machines. Voiceprint based biometric identifications is evolving as a new technique. Phonetic distance measurement is one such evolving technique and the cutting edge researchers are inspired to work on this technique to circumvent and overcome the above problem. This chapter covers a new speaker recognition model based on the pronunciation variability. The pronunciation variability is used to identify the voiceprint of the speakers. The Kullback-Leibler divergence relative entropy criterion is used for the speaker identification and verification. An adaptation model is designed for the unsupervised dynamic adaptation of the new pronunciation variants. The multi-layered code book memory using the modified vector quantization technique is designed to keep the word confusability low and ensures efficient retrieval of the pronunciation variants. The confusion matrix and performance metrics are used for performance evaluation of pronunciation classifier. The pronunciation classification error rate, OOV error rate and word error rate are used for evaluation.
