Frequency-time analysis approach to feature extraction for text independent speaker identification

摘要

This paper presents an alternative approach to Mel Frequency Cepstral Coefficient (MFCC) based method of feature extraction for robust text independent speaker identification. This work is focused to increase the identification accuracy without increasing the size and complexity of filter bank. The drive for this new feature extraction technique comes from a transformation which is based on the Nyquist filter bank consuructed using Gaussian filters. This new feature extraction technique has been compared with MFCC feature for different lengths of utterances. Experimental evaluation is carried out on MEPCO telephone speech database with 50 speakers using Gaussian Mixture Model (GMM). The proposed feature set performs significantly better than the MFCC feature set achieves 6% higher average accuracy compared to the MFCC feature set for utterance lengths of 20 seconds.