Feature extraction and selection for speech recognition

The major contributions of this thesis are: a new discriminative training algorithm, new discriminative feature selection and extraction algorithms, and a new image segmentation algorithm used for feature extraction from speech spectrogram.; In the first part of this thesis, a new misclassification measure and a discriminative training algorithm are proposed. The misclassification measure is a smooth representation of classification probability of error and can be made as close as possible to this probability by varying its parameters. The training algorithm indirectly minimizes the probability of error by minimizing the misclassification measure. A new discriminative training algorithm for speech segmentation based on another misclassification measure is also introduced.; In the second part of this thesis, a feature selection and a feature extraction algorithm are proposed. The proposed algorithms allow the dimensionality of feature space to be decreased, while trying to maintain a class separability measure. This measure is the misclassification measure of a classifier built in the higher dimensional space. The feature selection and extraction algorithms determine the maximum change in the misclassification measure (or indirectly the maximum loss in probability of correct classification) for the feature vectors presented in the lower dimensional space. The algorithms find the best subset of features and an optimum orthogonal linear mapping before applying feature selection that minimizes the maximum change in the misclassification measure.; In the third part of this thesis, several algorithms for feature extraction from speech spectrograms are proposed. Some of these algorithms first segment the spectrogram using a new self-organizing image segmentation algorithm. This algorithm segments the spectrogram into two classes of object and background, where pixels of each class have common characteristics. The algorithm iteratively minimizes a defined segmentation measure in the spectrogram image. Moreover, pixels with lower likelihood of belonging to object or background classes are adjusted less in each iteration, delaying their segmentation until more image information is available. The resulting features are the inputs to the proposed feature selection and extraction algorithms.; Some speaker independent isolated word speech recognition experiments are also carried out in this thesis which validate the proposed algorithms.