Speech adaptation using neural networks for connected digit recognition

The performance of speech recognizers is usually degraded when used in different environments due to varied channels, speech rate and so on. Retraining the recognizers demands a large amount of new data recorded under new environments. On the contrary adaptation can fit the characteristics of the new environments by using only a small amount of data. In this paper a neural network based adaptation was applied to enhance the performance of connected digit recognition system because of its ability of computing the nonlinear function. A baseline system was built on OGI Number corpus, which had 97.76% word accuracy and 85.19% sentence accuracy on itself. However when tested on the Australia English Telephone Speech Database the performance greatly decreased to 71.86% and 16.67% respectively. To avoid retraining the recognizer a feed-forward backpropagation network was used to fit the characteristics of new data, and reduced the error rate by 53% when combined with maximum likelihood linear regression.