Improvement of speech recognition results by a combination of systems

Abstract

The aim of this study is to suggest an algorithm that combines two speech recognition systems. These systems differ in the methods used in the feature extraction stage, but they have the same classifier Hidden Markov Model (HMM). The first system uses Mel-Frequency Cepstrum Coefficients (MFCC), the second one uses Linear Prediction Cepstrum Coefficients (LPCC), and the third system uses Perceptual Linear Predictive (PLP) features. The combination algorithm is applied separately on each couple of systems. The study is implemented on a data set that consists of the four voice commands: “shutdown”, “documents”, “restart”, and “net” pronounced by 33 people. In addition to the improvement of the speech recognition rate for isolated words, the study aimed to determine the most complementary couple of systems through studying two kinds of errors: simultaneous and dependent errors. The system depending on MFCC features provided the highest recognition rate with 85.44%. The results showed noticeable improvement of combined systems in comparison with the individual systems where combining MFCC & PLP provided the highest recognition rate with 93.44%.