抄録
This paper proposes a simple new model for the recognition of vowel sequences by a neural network using vocal tract area. Utilizing vocal tract area as input data reduces the number of input units and requisites computational time of learning compared with utilization of frequency spectrum. Practically, back-propagation rule is applied on three-layer structure, and suitable prediction of vocal tract area makes it possible to recognize speaker-independent vowel sequences by 39 units which is comparatively small as the number of total units.
