Abstract

Local features for any pattern recognition system are based on the information extracted locally. In this paper, a local feature extraction technique is developed, which captures the formant transition and voice onset/off set of a speaker. We named this technique as voice onset/offset local features (VOOLF). These features are extracted in the time spectrum domain by taking the moving average on the diagonal directions. These proposed features are compared with MFCC for speaker recognition system. The results showed that proposed technique perform better than the commonly used MFCC. The proposed method is able to capture the formant transitions and onset/offset of the speaker; hence this resulted in recognition rate higher than the other speech features.

Key words: Voice onset/offset features, local features, Speaker recognition system, Gaussian Mixture Model (GMM).