Isolated word recognition based on the adaptive neural network
J. S. Wang, Bernard Carlos Widrow · The Journal of the Acoustical Society of America · 1988
For the last several years, despite many new approaches (such as application of the Markov model and VQ) in speech recognition, the classical pattern matching using dynamic programming still prevails because it yields good recognition results. However, this DP-based algorithm requires a large amount of computation. It is difficult to meet the real-time requirement for large vocabularies. An alternative pattern recognition method called the adaptive neuron in adaptive control system used by B. Widrow in the early 1960s is used here to build an isolated word recognition system. A simple and efficient learning algorithm is presented for adaptively adjusting weight vectors to fit a certain word pattern. In order to capture the variations in speech, an increasing codebook was designed during the learning phase to allow increasingly more patterns; therefore, multiple reference patterns were adopted, giving the advantages of a statistical Markov model. The performance of this recognition system has been preliminarily tested on two common vocabularies including ten digits and the English alphabet. The recognition accuracy is 100% for digits after five trainings per word and 93% for the alphabet after ten trainings per word for speaker trained. Further experiments for a complete Chinese vocabulary are under progress.