Fusing wavelet and short-term features for speaker identification in noisy environment

Abstract

Effective Speaker Identification System (SIS) involves extracting features effectively. In this paper, we propose a feature extraction scheme based on wavelet analysis which is used along with short-term features. To overcome the drawbacks of Discrete Wavelet Transform (DWT), we propose to combine Stationary Wavelet Transform (SWT) with Mel-Frequency Cepstral Coefficient (MFCC) features. The combined features were used as inputs to K-nearest neighbors (Knn) classifier. The effectiveness of the proposed method is investigated for closed-set text-independent SIS in clean and noisy environments. The experimental results indicated that the proposed approach can achieve better identification rate performance with feature extraction using SWT rather than DWT.