Abstract
This paper presents a new system for spoken Japanese digits recognition by a neural network using vocal tract shapes. The vocal tract shape is a suitable parameter for synthesis or recognition. The vocal tract shapes are used for the neural network as input data. We first propose a simple method by which the vocal tract shape is directly estimated from speech waves. A three-layered neural network is used in our recogni tionsystem. The network learning algorithms utilized here are conjugate gradient (CG) algorithm and backpropagation (BP) algorithm. Finally, we show the recognition results to prove the effectiveness of our method, and we show that the CG algorithm has several advantages compared to the BP algorithm.
