Effect of Feature Extraction Techniques on the Performance of Speaker Identification

Abstract

In this paper, the effect of features extracted on the performance of speaker identification engine is investigated. Vector Quantization (VQ) is implemented and used as identification engine. Three type of speech features, Mel Frequency Cepstral Coefficients (MFCC), Perceptual Linear Predictive (PLP), and Relative Spectral Technique- Perceptual Linear Predictive (RASTA-PLP) are extracted and used for the classification problem. One word per speaker is used within the train phase and the identification rate is calculated for each feature extraction technique. The calculation is repeated using various word of different spoken time, and the paper specifies the feature extraction technique that fits with the Vector Quantization (VQ) recognition engine. 