Golik A. Compliance distances and clustering techniques for vector and matrix feature vectors and their usage in speech and gesture recognition

The dissertation is devoted to development of models, methods and algorithms for clustering of speech and gestures. Different approaches to formation of feature vectors in vector and matrix form are covered. Development of the Moore-Penrose pseudoinverse theory, in particular, formation of compliance distance based on immersion into appropriate subspace (or hyperplane) in space of features is suggested. Techniques of standardization of spectrogram, image of gesture and geometrical characteristic of gesture, smoothing a matrix of spectrogram using SVD-decomposition are given. Two recognition systems were developed and implemented: recognition of speech of one speaker on a limited set of words and recognition of a tactile sign language with efficiency 93% and 86% correspondingly.