Text-independent pronunciation evaluation based on phone-level Gaussian classifier

This paper presents a novel text-independent pronunciation evaluation method. The method aims at testing the mandarin Putonghua pronunciation proficiency of the dialectal speakers from every area of China. In the proposed method, 5 pronunciation proficiency levels are assigned for each phone; accordingly, 5 Single Gaussian classifiers are created to represent the 5 levels for a phone. 4 phone-level features are described and used in the classifiers. In the testing step, phone-level output probabilities are combined to get the sentence-level likelihood.