Evaluation of Cepstral Features of Speech for Person Identification System Under Noisy Environment

Abstract

Robust feature extraction techniques play an important role in speaker recognition system. Four speech feature extraction techniques such as Mel-Frequency Cepstral Coefficient (MFCC), Linear Prediction Cepstrum Coefficient (LPCC), Perceptual Linear Predictive (PLP), and Wavelet Cepstral Coefficient (WCC) techniques are analyzed for extracting speaker-specific information. The design of WCC method is done for this work. Hidden Markov Model (HMM) is used to model each speaker from the speaker-specific speech features. The conventional Person Identification System (PIS) is normally employed in an environment where the background noise is unavoidable. To simulate such environment, an additive white Gaussian noise of different SNRs is added with a studio quality speech data. Evaluation of PIS is performed using the Hidden Markov Toolkit (HTK). Multiple experiments are performed. Acoustic modeling of speaker and evaluation is done for clean and noisy environment. The experiment results indicate that 100% accuracy for text-independent PIS in a clean environment. Furthermore, it is observed that MFCC is proven to be better noise robust than PLP and LPC. It is also noted that dynamic features such as delta and acceleration features are combined with static features improve the performance of the PIS in noisy environment.