Algorithm and method for quantitative assessment of the speech signals similarity

Abstract

The paper proposes a method to solve the task of automated quantitative assessment of the syllable pronunciation. This quantitative assessment is used to evaluate the speech quality during speech rehabilitation. An algorithm for quantifying the similarity of two audio signals of different lengths is presented. The algorithm uses a hybrid match measure. The hybrid measure is based on calculation of three metrics (DTWdistance, correlation coefficient and Minkowski metric) and using a fuzzy classifier as a mechanism for combining the calculated values. The average number of coincidences of estimates by the proposed algorithm and estimates by the previously applied method is 83%. A method for quantifying the similarity of speech signals using several reference signals is proposed. The method allows to consider the variability of speech and the individual characteristics of the phoneme’s pronunciation. This is achieved by using records of the patient's preoperative speech as reference signals.