Quantitative Analysis of Feature Extraction Techniques for Isolated Word Recognition

Abstract

Isolated word recognition has been a subject of research since the 1940s when speech recognition technology was in a nascent stage. Recurrent neural networks and deep feed-forward networks are currently being explored by researchers to increase the efficiency of the speech recognition systems. However, probabilistic techniques like Gaussian Mixture Model (GMM) and Hidden Markov Model (HMM) have been the state-of-the-art since long. This paper performs a quantitative analysis of feature extraction techniques for isolated word recognition to provide better insights in improving the efficiency of the system. In this regard, a basic architecture of the word recognizer system has been modelled and it has been observed that Mel Frequency Cepstrum Coefficients (MFCC) in combination with Delta and Delta-Delta parameters have 92.4% accuracy for a sufficiently large dataset. Also MFCC features, appended with Delta parameters have 87.0% accuracy which is 36.4% higher than that of Short Time Fourier Transform (STFT) features. The feature extraction techniques have been classified by a Gaussian Mixture Model- Hidden Markov Model (GMM-HMM) classifier. This paper also studies the effect of varying data size and number of features on the recognition model towards efficient word recognition. These recognizers may then be used as a building block for mispronunciation detection in a Computer Aided Pronunciation Training (CAPT) system.