Automatic pronunciation evaluation and classification

Pronunciation evaluation is an important module of every spoken language evaluation system. Automatic evaluation of quality of pronunciation that can mimic the performance of human assessors is a difficult task as human assessment accounts for several nuances of pronunciation including vowel substitutions and quality of consonants. This paper presents a novel approach that combines the knowledge of human assessment and the knowledge of the behaviour of automatic speech recognition systems to develop features for pronunciation evaluation. Instead of presenting the correlation of the proposed features with human assessment, the paper presents sentence-level classification accuracies which can directly be used in real-life applications. Inter-human and intra-human agreements, which are indicative of human subjectivity, are also presented. The trends in confusions among humans scores and automatic scores are compared as the number of classification classes is varied.