Discriminative Training using Heterogeneous Feature Vector for Hindi Automatic Speech Recognition System

Training and testing are the two phases that are used in statistical approach of designing an automatic speech recognition (ASR) system. The training phase includes parameterization of input speech signal and acoustic modeling of speech features. The paper proposes discriminative training of hidden Markov Model (HMM) that uses heterogeneous feature vector for continuous Hindi ASR system. A linear interpolation of mel frequency cepstral coefficients (MFCC) and perceptual linear prediction (PLP) is used to generate heterogeneous feature streams. The implemented work uses maximum mutual information estimation (MMIE) and minimum phone error (MPE) discriminative techniques for acoustic model training. The results show that MF-PLP parameterization with MPE discriminative techniques combination outperforms the other feature extraction and discriminative combination.