Automated assessment of English fundamental frequency contours for non-native speakers from China and India

• Published in: The Journal of the Acoustical Society of America 2013-11, Vol.134 (5_Supplement), p.4068-4068
• Main Authors: Evanini, Keelan, Wang, Xinhao
• Format: Article
• Language: English
• ISSN: 0001-4966; 1520-8524
• DOI: 10.1121/1.4830852

This study investigates the F0 contours produced by non-native speakers of English from two different countries. Speakers from China (N = 202; L1 Mandarin) and India (N = 230; multiple L1s) read a paragraph out loud in the context of an assessment of non-native English speaking proficiency; in addition, a control set of native speakers (N = 85) read the same paragraph. The words and phonemes in all of the responses were provided with time stamps using forced alignment with the prompt text, and mean normalized F0 values were extracted for each word from each speaker. A model F0 template for the paragraph was generated by taking the mean word-level F0 value across all native speakers, and the correlation between the word-level F0 contour and this template was calculated for each non-native speaker (as in Schwanenflugel et al. 2004). This correlation was used as a feature for assessing F0 contours and was correlated with expert human scores of English proficiency. The results demonstrated that the feature performed somewhat better for the speakers from India (r = 0.390) than the speakers from China (r = 0.326). Furthermore, the correlations improved when function words were removed from the analysis and only content words were considered (r = 0.396 and r = 0.346).