A Feature Selection Method in Spectro-Temporal Domain Based on Gaussian Mixture Models

Spectra-temporal representation of speech is considered as one of the leading speech representation approaches in speech recognition systems in recent years. This representation is suffered from high dimensionality of the features space which makes this domain unusable in practical speech recognition systems. In this paper, a new method of feature selection is proposed in the spectro-temporal domain. In this method, clustering techniques are applied to spectro-temporal domain to reduce the dimensions of the features space. In the proposed approach, spectro-temporal space is clustered based on Gaussian Mixture Models (GMMs). The mean vectors and covariance matrices elements of the clusters are considered as a part of the feature vector of the frame. The tests were conducted for new feature vectors on voiced stops (/b/, /d/, /g/) classification of the TIMIT database. Using the new feature vectors, the results were improved to 70.45% which is 7.95% higher than last best results.