Tied Mixture Modeling in Hindi Speech Recognition System

Abstract

The goal of automatic speech recognition (ASR) is to accurately and efficiently convert a speech signal into a text message independent of the device, speaker or environment. In ASR, the speech signal is captured and parameterized at front-end and evaluated at back-end using the Gaussian mixture hidden Markov model (HMM). In statistical modeling, to handle the large number of HMM state parameters and to minimize the computation overhead, similar states are tied. In this paper we present a scheme to find the degree of mixture tying that is best suited for the small amount of training data, usually available for Indian languages. In our proposed approach, perceptual linear prediction (PLP) combined with Heteroscedastic linear discriminant analysis (HLDA) was used for feature extraction. All the experiments were conducted in general field conditions and in context of Indian languages, specifically Hindi, and for Indian speaking style. © Springer-Verlag Berlin Heidelberg 2010.