Third International Conference on Spoken Language Processing (ICSLP 94)
In this paper, we present a method of generating articulatory trajectories from formant transitions. A multilayer feedforward neural network controls a vocal tract model on the basis of its first three resonances (the formants) in order to compute 8 articulatory parameters (the cross sectional areas of 7 regions of the Distinctive Regions Model and the vocal tract length). 4 learning modes have been studied and compared: autonomous, supervised, hybrid and supervised-hybrid. After training in one of the 4 modes, the neural network is able to generate articulatory configurations. Each formant triplet time sample is successively introduced into the network. For each triplet, a local optimisation is performed by the neural network in an autonomous learning mode. Several articulatory trajectories are presented.
Bibliographic reference. George, Martine / Jospa, Paul / Soquet, Alain (1994): "Articulatory trajectories generated by the control of the vocal tract by a neural network", In ICSLP-1994, 583-586.