Two neural networks, multi-layer perceptron(MLP) and neural prediction model (NPM), are trained separately for speaker independent isolated word recognition, and later combined by another MLP for improved performance. The first MLP classifies the word as a spectro-temporal pattern, while the NPM relies on dynamic nature of the speech signal. To reduce misclassifications the second MLP is trained to extract decision rules and serve as an intelligent "judge" for disputes between two classifiers. The second MLP accepts input values from two preceding neural network classifiers, and provides desired classification. The combined network utilizes both spectrotemporal pattern and dynamic nature of speech, and demonstrates better recognition rate. It shows less sensitivity to the choice of training data.
