Relative Contribution of Vowel Quality and Duration to Native Language Identification in Foreign-Accented English

Automatic identification of a speaker's native language back-ground may have forensic applications. This paper explores the feasibility of automatic identification of the native language background of a foreign speaker of English, using phonetically interpretable measurements. The production of the ten mono-phthongs of (American) English by Dutch, Mandarin Chinese and American speakers was used as a test case. Vowel formants F1 (corresponding to articulatory vowel height), F2 (capturing vowel backness and lip rounding) and vowel duration were extracted. Clearly different duration and patterning of the vowels in the vowel space were seen. Automatic classification of the speaker's native language was 90 percent correct when all acoustic parameters were used as predictors. Language identification was slightly poorer when only formant data were used (85% correct) and substantially poorer -- but much better than chance -- when only vowel duration was used (60% correct). We conclude that vowel duration provides a weaker cue to foreign-accent identification in English than the spectral properties but that the combination of both information sources yields the best results.