The Formal description and experimental investigation of context-based features for the classification of segmentation instants of speech signal
ENIn this work, the problem of data driven speech signal segmentation is addressed. Approach of the Work is based on the symbolic machine learning (rule induction) techniques and on the idea of context-based feature set which is used to describe candidate segment boundaries. Symbolic machine learning is used for automatic generation of Boolean functions discriminating true and false segment boundaries. Objectives of this work are: 1) to present a formal description of context-based features which is built not only on the local properties of a time instant corresponding to the candidate segmentation mark, but comprises properties and relationships of other neighboring time instants as well; 2) to demonstrate the application of these context-based features for segmenting speech signals, using rule learning algorithm RIPPER-k.
