Mark Hasegawa-Johnson
Res. Lab. of Electronics, MIT, 36-549, Cambridge, MA 02139
Given an accurate burst location, four automatic measurements of the burst spectral shape and three semiautomatic formant onset measurements were used to correctly classify place of articulation of 90% of a corpus of 124 stops. At least 20 releases of each of the 6 English stops, in unconstrained prevocalic, preglide, or (for velar and labial stops only) pre-/r/ context, were chosen at random from the TIMIT corpus. Burst locations were marked by hand. The values of F2 and F3 20 ms later, and of (Delta)F2 between 20 and 50 ms later, were measured automatically, and 18% of the tokens were corrected by hand. Burst compactness, peak frequency, and mid- and high-frequency amplitude were measured automatically. In separate analyses of variance, all seven measures showed significant place effects (p<0.001). In three pairwise linear discriminants, the burst measures alone were able to classify place with 85% accuracy, showing that automatic measurements can reproduce the result of Blumstein and Stevens [J. Acoust. Soc. Am. 64 (1978)]. With seven measures, the error rate was only 10%. Most errors could be attributed to a strong labial burst and rapid formant transitions in /i/ or /r/ context, or coloration of an alveolar burst by back-cavity coupling. [Supported by a grant from NIDCD.]