SVM-based evaluation of Thai tone imitations by Thai-naïve Mandarin and Vietnamese speakers

Native listener judgements and acoustic comparisons are sensitive to deviations between non-native speech and native productions, but both have drawbacks and are inefficient for evaluating large databases. To probe whether Support Vector Machines (SVM) might offer an efficient alternative, we used three SVM models trained with native Thai lexical tones to eval-uate new native stimuli and non-native imitations by Mandarin and Vietnamese speakers. The optimal SVM model categorized native tones accurately but showed lower accuracy with non-native imitations, like native judges do, thus confirming its sensitivity to deviations from native productions. Thai falling tone imitations yielded the lowest classification accuracy, indicating that both groups' imitations were constrained by their native falling tones. Thai rising tones were better recognized for Viet-namese than Mandarin imitators, reflecting differences between their native rising tones. Thus, SVM modeling may provide an effective alternative to traditional perceptual- or acoustic-based evaluations of non-native speech.