In addition to measuring job candidates spoken English proficiency, quantifying the degree of accentedness may help companies assign employees to appropriate job categories, or identify employees who could benefit from additional speech training. In this paper, we discuss methods for automatic accent quantification of Indian English speakers. Similar to techniques used in speaker recognition, we used Gaussian mixture models (GMMs) for the modeling of accent spectral characteristics in different groups of subjects. Computationally, we verified that certain consonants in Indian English have more discriminative power than others in quantifying an Indian accent. As a result, we propose the idea of using GMMs to model only certain phonemes with high predictive power. By combining features from GMMs with others, we achieved a human-machine correlation coefficient of 0.84 at the participant level. The results validate the use of new proposed methods to quantify accents automatically.
Bibliographic reference. Cheng, Jian / Bojja, Nikhil / Chen, Xin (2013): "Automatic accent quantification of indian speakers of English", In INTERSPEECH-2013, 2574-2578.