Feature Extraction Using LPC-Residual and MelFrequency Cepstral Coefficients in Forensic Speaker Recognition

Abstract

Abstract-In this paper, we investigated the form to improve the performance in the recognition, involved at the forensic area. To improve, we use Linear Predictive Coding (LPC) and it residual; it was compared with Mel Frequency Cepstral Coefficients (MFCC). The classification technique was Gaussian Mixture Model (GMM). The collection data is in Spanish language, using spontaneous speech from 37 male speaker of Mexican Spanish, we have three recordings, between each recording exist 3 week and one month of separation respectively, this allows us in real condition work, we use non contemporaneous recording and scarcity of data to training and testing the performance of the forensic recognition task. Two conclusions can be drawn from the results, the first, MFCC has better performance with long recording, LPC-residual has better performance with short recording. Index Terms-Forensic speaker recognition (FSR), gaussian mixture model (GMM), linear predictive coding-residual (LPC-residual), mel frequency cepstral coefficients (MFCC).