Robust Text-independent Speaker Identification in a Time-varying Noisy Environment

Abstract

Practical speaker recognition systems are often subject to noise or distortions within the input speech which degrades performance. In this paper, we proposed a new mel-frequency cepstral coefficients (MFCC) based speaker identification system with Vector Quantization (VQ) modeling technique. It integrates a hearing masking effect based masker and a group of dozen triflers into traditional MFCC feature extraction for robust speaker identification. The masker can decrease the influence of noise signal to the speech signal, and improve the recognition rate. The mixture of triflers can enhance high-frequency calculation accuracy. A purposeful voice samples database are collected under an unconstrained indoor environment for a month. The texts to be spoken are also unconstrained. The proposed method is evaluated with the voice samples database, and its recognition rates remain over 93% under different experiment condition. The experiments results show that the proposed speaker identification system has good accuracy and robustness to the unconstrained noisy condition and text-independent.