Dept. Signal, Unite Associee au CNRS, ENST, 46 rue Barrault, 75634 Paris Cedex 13 France
Univ. of Texas, Austin, TX
An algorithm is described that takes an arbitrary VT area function (normalized to a length of 18 cm), calculates its formant frequencies and their ``sensitivity functions,'' and uses those functions to induce VT shape changes that increase or decrease a given formant frequency. The computations are performed recursively, and for each formant separately, until extreme values are established. Irrespective of the starting configuration, the procedure converges on a given set of formant ranges that closely match observed spoken values. Thus in its efforts to enhance ``acoustic contrast'' (increase formant ranges), the algorithm gradually ``rediscovers'' the human vowel triangle. Final VT perturbations show isomorphism with the ``distinctive regions'' of the DRM model [Carre and Mrayati (1992)]. One significant consequence of that finding is that an input configuration lacking a human ``pharynx'' will eventually be transformed into a shape that does have such a cavity. These results are compatible with the role accorded to ``acoustic contrast'' in the formation of vowel systems. Also they throw some new light on the old issue whether the evolution of VT tract morphology was (in part) an acoustically driven process. [sup a)]This work was performed during a stay at Dept. of Linguist., Univ. of Texas at Austin.