Refinement of HMM Model Parameters for Punjabi Automatic Speech Recognition (PASR) System

ABSTRACT An automatic speech recognition system follows an approach of pattern matching, which consists of a training phase and testing phase. Despite advancement in training phase, the performance of the acoustic model is adverse while adopting the statistical technique like hidden Markov model (HMM). However, HMM-based speech system faces high computational complexity and becomes challenging to provide accuracy during isolated Punjabi lexicon. As the corpus of the system increases, the complexity of training phase will also increase drastically. The redundancy and confusion occurred between feature distributions in training phase of the system. This paper proposes an approach for the generation of HMM parameters using two hybrid classifiers such as GA+HMM and DE+HMM. The proposed technique focuses on refinement of processed feature vectors after calculating its mean and variance. The refined parameters are further employed in the generation of HMM parameters that help in reduction of training complexity of the system. The proposed techniques are compared with an existing technique such as HMM on benchmark database and self-developed corpus in clean, noisy, and real-time environments. The results show the performance improvement in pattern matching of spoken utterance when demonstrated on large vocabulary isolated Punjabi lexicons.