Abstract:

A large stimulus continuum (972 stimuli in an F1--F2-F3 space) was categorized by both Finnish and English listeners. Further results of logistic regression modeling of listeners' responses using formant frequencies and/or a spectral shape representation (warped-frequency LPC cepstral coefficients) are reported. Cross-validation model selection methods [J. Shao, J. Am. Stat. Assoc. 88, 486-494 (1993)] indicate that a combination of the two parameter sets (formant and cepstral) is superior to either alone. However, for both languages the best prediction scores were obtained from a purely formant-based analysis that included squares and cross products of frequency values, thereby allowing for quadratic category boundaries in the formant space. Analysis to date provides no evidence that spectral-shape information (versus formant frequencies alone) plays a role in listeners' perception of these stimuli in either Finnish or English. Results of additional analyses will be presented at the meeting. [Work supported by SSHRC.]