Machine Learning-Based Approach to Transcribing Language Units

The article addresses the task of automated transcribing of speech signals within the framework of pattern recognition theory. To tackle this task, an adaptation of the decision tree method is proposed in the form of a sequence of tasks of multi-element statistical classification of phonemes in speech signals. The algorithm is controlled by two threshold values that determine the permissible divergence between the investigated and reference signals characterizing the same phonemes in terms of relative entropy metric, and restrict the maximum duration of phonation for the corresponding speech element. Experiments demonstrated both the functionality of the proposed algorithm and its adaptability, provided by the variation of the mentioned threshold values.