Fitness Evaluation of Gaussian Mixtures in Hindi Speech Recognition System

Speech recognition (i.e. speech to text conversion)and speech synthesis (i.e. text to speech conversion)are two main operations performed in human computer interaction through natural language conversational interface. This paper presents a novel approach for modeling and designing of a Hindi speech recognition system, by using Perceptual Linear Prediction (PLPRASTA)for feature extraction and Gaussian Mixture Model (GMM) for statistical pattern classification. Experimental results show that only 4 Gaussian mixtures yield optimal performance in the context of small databases available for Indian languages which have been used to train the Hidden Markov model (HMM). The results also illustrate that the approach presented in this paper, not only outperform the traditional model for small vocabulary in typical field conditions but can also be implemented efficiently in embedded systems.