A comparative study of using different speech parameters in the design of a discrete hidden Markov model

Linear predictive coding (LPC)-based cepstral coefficients are the most widely used parameters in the design of modern speech recognizers. The authors compared useful features, such as the LPC constants themselves, filter bank analysis coefficients, and short-time Fourier transform coefficients, with respect to recognizer accuracy. The recognizer studied is based on the discrete hidden Markov model (HMM). For fairness of comparison, all of the different feature vectors were of the same dimensionality. Also, to compare computational requirements for the various features, the size of a frame in the typical frame-by-frame discrete analysis and the frame sampling rate were kept similar