An artificial neuron approach to speech processing based on a hidden Markov model

The paper discusses possible designs for a hidden Markov model using artificial neural network techniques. Applications are speech generation and speech recognition. The designs could be the basis for a computer program to realise an HMM; or they could be inspiration for a special purpose piece of hardware. The designs have two parts. The first determines the time sequence of hidden states. Although the design uses mainly standard logic components, the behaviour will be similar to a backpropagation layer of ANNs. The second part consists of one or more layers of feedforward AN processing elements in a standard backpropagation network with final outputs coded speech data. For training an HMM recognizer, the normal approach is to use the Viterbi Algorithm followed by fine tuning using the Baum Welch re-estimation algorithm. The ANN-HMM can be trained this way, but it can also be trained by the backpropagation algorithm, BP. The latter ties very directly to the unknown weights in the ANN circuits and may facilitate topological design.