LINEAR TRAJECTORY MODELS INCORPORATING PREPROCESSING PARAMETERS FOR SPEECH RECOGNITION

In this letter, we investigate the interactions of front-end feature e xtraction and back-end classification techniques in nonstationary stat e hidden Markov model (NSHMM) based speech recognition. The proposed m odel aims at finding an optimal linear transformation on the mel-warpe d discrete Fourier tranform (DFT) features according to the minimum cl assification error (MCE) criterion. This linear transformation, along with the NSHMM parameters, are automatically trained using the gradien t descent method. An error rate reduction of 8% is obtained on a stand ard 39-class TIMIT phone classification task in comparison with the MC E-trained NSHMM using conventional preprocessing techniques.