Comparative study between MFCC and LP-Mel based front-ends for noisy speech recognition using HMM

Abstract

Since the parameterization in the perceptually relevant aspects of short-term speech spectra in ASR front-end is advantageous for speech recognition, such as Mel-LPC, LPC-Mel, MFCC etc., in this paper, MFCC and LP-Mel based front-ends have been designed for automatic speech recognition (ASR). The speech classifier of the developed ASR is based on Hidden Markov Model (HMM) as it can successfully cope with acoustic variation and lack of word boundaries of speech signal. The performance of the developed system has been evaluated on test set A of Aurora-2 database both for MFCC and LP-Mel based front-ends. It has been found that the MFCC based front-end is more effective for noise type subway, babble, car and exhibition. The average word accuracy for MFCC has been found to be 59.21%, while for LPC-Mel, it has been 54.45%.