Deep neural network based supervised speech segregation generalizes to novel noises through large-scale training

Deep neural network (DNN) based supervised speech segregation has been successful in improving human speech intelligibility in noise, especially when DNN is trained and tested on the same noise type. A simple and effective way for improving generalization is to train with multiple noises. This letter demonstrates that by training with a large number of different noises, the objective intelligibility results of DNN based supervised speech segregation on novel noises can match or even outperform those on trained noises. This demonstration has an important implication that improving human speech intelligibility in unknown noisy environments is potentially achievable. OSU Dept. of Computer Science and Engineering Technical Report #02, 2015