Using Gaussian Mixtures for Hindi Speech Recognition System
Rajesh Kumar Aggarwal, Mayank Dave · 2011
The goal of automatic speech recognition (ASR) system is to accurately and efficiently convert a speech signal into a text message independent of the device, speaker or the environment. In general the speech signal is captured and pre-processed at front-end for feature extraction and evaluated at back-end using the Gaussian mixture hidden Markov model. In this statistical approach since the evaluation of Gaussian likelihoods dominate the total computational load, the appropriate selection of Gaussian mixtures is very important depending upon the amount of training data. As the small databases are available to train the Indian languages ASR system, the higher range of Gaussian mixtures (i.e. 64 and above), normally used for European languages, cannot be applied for them. This paper reviews the statistical framework and presents an iterative procedure to select an optimum number of Gaussian mixtures that exhibits maximum accuracy in the context of Hindi speech recognition system.