Acoustic Modeling of Speech Units Using GVQ Probability Density Representation
Abstract
In this paper, a novel probability distribution representation method has been proposed for acoustic speech segments. This representation is applicable to various acoustic models. In this method, the probability density of each feature vector is calculated by a selected Gaussian distribution, which is compatible with environmental conditions. This Gaussian distribution is selected among pre-estimated Gaussian clusters of each phoneme. In this regard, the required training and recognition algorithms are developed and analyzed in this paper. It is shown that the training time is reduced drastically while the recognition rate has remained unchanged. In addition, the GVQ method makes an appropriate analytical framework for Gaussian selection approaches to speed up the recognition phase.
Keywords
automatic speech recognition, acoustic modeling of speech, parametric stochastic representation, Gaussian mixture models, vector quantization, Gaussian vector quantization
