15th International Congress of Phonetic Sciences (ICPhS-15)Barcelona, Spain |
We report some results of modeling the categorization of a three-formant synthetic vowel continuum by speakers of English and of Finnish. We focus here on assessing the relative merits of initial representations based on a standard 3-dimensional formant-frequency space compared to those based on 1) a 2-dimensional F1 by F2- prime space, and on 2) Hermansky's PLP5. PLP is a 5- dimensional representation of spectral shape that manifests an integration of closely spaced formants in a manner suggestive of F2-prime. Linear logistic models suggest that PLP5 modestly outperforms 3 formants, which in turn substantially outperforms F1 by F2-prime. However, if quadratic rather than linear methods are allowed, the 3 formant space provides a better account of listeners' response patterns than do other models of comparable complexity. This result is confirmed using modern model selection techniques.
Bibliographic reference. Nearey, Terrance M. / Kiefte, Michael (2003): "Comparison of several proposed perceptual representations of vowel spectra", In ICPhS-15, 1005-1008.