Evaluation of a network-based isolated digit recognizer using the TI multi-dialect database

This paper describes three sets of experiments based on the adult-talker, isolated digits portion of Texas Instruments' multidialect database which are aimed at evaluating the cross-dialect performance of our network-based digit recognizer. The first two sets of experiments examine system performance as a function of the dialectical diversity of the talkers used for system training. The third set examines recognition scores for a system which makes use of networks separately trained on male and female talkers. Recognition accuracies of 96% to 99% were obtained when the dialectical diversities of the talkers used for training and testing were comparable. Performance improved by 0.5%, on average, when separate male/female networks were used.