Speech Assessment Based on Entropy and Similarity Measures

Abstract

Identifying similarities between audio signals is an important goal for the Speech Recognition Systems. This paper proposes a new method for the assessment of the similarity between two audio signals. Based on the entropy theory, the audio signals are segmented and compared on the basis of the information value carried by each segment: this value is obtained considering an alphabet derived from the trend lines of each audio signal. It is an experimental method that permits to represent and analyse the results through an isometric diagram, for a better interpretation of them.