Frequency Cepstral Coefficient and Learning Vector Quantization Method for Optimization of Human Voice Recognition System

Authors

  • Cecep Sandi, Ari Octa Pratama Riadi, Fathul Khobir, Agung Laksono, Ari Purno Wahyu W

Abstract

Speaker verification is the process of verifying a speaker, where the identity of the speaker is previously known based on the data that has been inputted. To be able to perform speaker verification, voice data will go through the process of extracting voice characteristics to obtain the information contained in the voice data. In previous studies, the Mel Frequency Cepstral Coefficients (MFCC) method has the highest accuracy rate with a recognition rate of 85.3% and the fastest feature extraction time compared to other feature extraction methods, so the MFCC method is a good method for feature extraction in speech recognition. As a classifier, the Learning Vector Quantization (LVQ) neural network is used. LVQ is a method for conducting learning or training at a supervised competitive layer. Before the voice data is processed, first the voice data is extracted using the MFCC method. The results of feature extraction using the MFCC method are used for the matching process, where this matching process will compare the results of feature extraction from the test data with the results of feature extraction from the training data contained in the database. Data from feature extraction were classified using LVQ. LVQ performs learning on vectors from the results of the MFCC filter. The sound matching process can be done by measuring the closest distance using the Euclidean distance to find out how similar the sound from the test data is to the pattern data in the knowledge base in the database  

Published

2020-11-01

Issue

Section

Articles