An ASR system using MFCC and VQ/GMM with emphasis on environmental dependency

Abstract

Automatic speaker recognition (ASR), also known as voice biometric recognition, remains very popular research area over six decades. Among all the acoustic features that are used in ASR, Mel-frequency Cepstral Coefficients (MFCCs) and Gammatone Frequency Cepstral Coefficients (GFCCs) are the most popular ones. However to make ASR environment independent, Relative Spectral Amplitude (RSATA) filtering techniques before feature extraction and feature, model, and score (in classification step) domain normalization techniques are applied. The techniques for modeling/classification that are used in present days are Vector Quantization (VQ), Support Vector Machine (SVM), Gaussian Mixture Models (GMMs), Hidden Markov Model (HMM), Artificial Neural Network (ANN), Deep Neural Network (DNN). In this paper we cite our experimental results upon three databases, namely, Hyke-2011, ELSDSR and IITG-MV SR Phase-I, based on MFCCs and VQ/GMM where Maximum Log-Likelihood (MLL) scoring technique is used for the recognition of speakers. The experimental results in the environmental mismatch condition for the IITG-MV SR Phase I & II databases are provided with explanation of accuracy degradation.