ASA 127th Meeting M.I.T. 1994 June 6-10

2pSP7. Recovering task-dynamic parameters from formant frequency trajectories: Results using computer ``babbling'' to form indexed files.

Richard S. McGowan

Haskins Labs., 270 Crown St., New Haven, CT 06511

Two indexed files of acoustic-articulatory pairs were created: one with 11 156 entries and another about ten times as large. Each individual entry was generated by running the Haskins articulatory synthesizer with randomly generated task-dynamic parameters. Each individual consisted of acoustic data in the form of the first three formant frequency trajectories and articulatory data consisting of the corresponding task-dynamic parameters. Each individual was indexed with keys denoting the direction of change in each formant frequency and the amount of change within 10% intervals. A genetic algorithm was used to recover the task-dynamic parameters from the formant frequency trajectories of four synthesized utterances. One recovery condition had a random initial population and the others used indexed files. Two factors were studied: the size of the indexed file and the criterion for an individual to be including into the initial population. One criterion was that the direction of formant change had to match that of the data, and the other was that the amount of formant change also had to match. The results were that using the larger indexed file provided faster convergence, but the results on matching criterion were ambiguous. [Work supported by NIH Grant No. DC-01247.]