Learning the time-delay characteristics in a neural network

Neural networks with time delays have been used successfully for speech recognition, but most of these networks have fixed delays and a training algorithm is used only for learning the pattern of connections. A learning algorithm that can also be used to train the characteristics of time delays is introduced. The learning algorithm minimizes a mutual discrimination error measure. Networks where the connection strengths and delay parameters are learned can be used to recognize temporal patterns (words) of arbitrary lengths. The feasibility of this algorithm is shown on a network that has been used previously with fixed delays for digit recognition. Improvement in recognition accuracy due to training the delay characteristics is demonstrated using a multispeaker digit database.<<ETX>>