Performance of Speaker Recognition System in Mismatching Speaking Style

Abstract

Analysis of speaker recognition system under different speaking style and mismatch conditions is presented. Generally, in text independent speaker recognition system, the way of speaking style used remains the same in training and testing phase. Whisper speech is generally used quietly to convey secret information or to avoid disturbing others in quiet place, so hearing of speech is limited to nearby listener only. Source of excitation and vocal tractsystem contains speaker specific information. Vocal folds do not vibrate in whisper speech so source excitation related information is not present. Fast speech is a tendency to speak rapidly, as if motivated by urgency unobvious to listener. The CHAINS speech corpus: CHAracterizing INdividual Speakers database in whisper, solo and fast speaking style is used for performing experiments with Gaussian Mixture Modeling- Universal Background Modeling (GMM-UBM) approach with the most widely used feature Mel Frequency Cepstral Coefficients (MFCC). The mismatch condition in training and testing speaking style is observed from the experiments and the performance degradation is found to be high when mismatch of train-test condition is considered.

Cite this Article

Pinky J. Brahmbhatt1, K.G. Maradia, Performance of Speaker Recognition System in Mismatching Speaking Style. Journal of Artificial Intelligence Research & Advances. 2018; 5(3): 58–65p.