Evaluación comparativa de sistemas de reconocimiento de locutor basados en los algoritmos LPC, CC y MFCC

Abstract

This document proposes the evaluation of speaker recognition systems based on the LPC (Linear Predicting Coding), CC (Cepstral Coefficients) and MFCC (Mel Frequency Cepstral Coefficients) algorithms, used in the extraction of voice parameters. The evaluation, following an experimental quantitative methodology, consists of determining the change in performance when the input signal is exposed to different noise conditions (crowd and Gaussian noise), namely, at different levels of SNR, comparing the verification results for 2 speakers. Although all the systems decrease their performance in noisy environments, each one possesses intrinsically a certain level of robustness. This evaluation will serve as a reference in the construction of speaker recognition systems, which include voice enhancement systems to reduce noise.