任意音声生成のためのニューラルネットワークにおける学習の高速化

概要

論文の詳細を見る
The authors have proposed a new neural network model and learning algorithm for speech production. This method has three features: (1) removal of singular points of back propagation (BP) errors; (2) a spread pattern information (SI) learning algorithm; (3) a cooperative learning (CL) algorithm. In this paper, the convergence problem of the neural network is discussed in the case where the feature (1) is applied to the SI method (2). First, from speech experiments it is shown that activities of the hidden units, which have two saturated states, change quickly from a saturated state to the other saturated one, and the rate in the saturated states is more than 80%. When the unit is in a saturated state, the amount of BP errors is small even though the output from the neural network is far from the target. This is why the convergence of learning is slow. Second, it is derived theoretically that removing the singular points of BP errors (feature (1)) make the system stable because this method shrinks the nonconvergent region. Finally, the performance of the neural network is evaluated in speech experiments. As a result of these experiments, the convergence of the proposed method is about twenty times faster than that of the original BP method.
東海大学の論文
1994-03-30