Generalization of deep neural-network based noise reduction to accented-English speecha)

The current study involved the ability of deep-learning based noise reduction to generalize to accented-English speech. The deep neural network (DNN) was trained using a large number of different speakers (and different noises) of primarily standard native American English. This model was used to remove background noise from speech produced by 88 speakers having 9 different native-language backgrounds, who spoke English as a second language. The resulting speech was characterized by accents strong enough to substantially affect intelligibility. Objective measures of intelligibility, which reflect acoustic accuracy of the signal following noise reduction, indicated that the DNN generalized to accented speech well. Human-subjects intelligibility testing supported this conclusion. These results are promising in terms of the ability to deploy DNN-based noise reduction in hearing devices to broad populations, including those who receive speech from non-native speakers. Additional objective measures of speech quality indicated that the DNN led to slightly higher subjective sound quality for speech that had lesser accents and was more similar to the speech used for model training. This result suggests that expanding the training set to include accented speech has the potential to improve subjective sound quality of the processed signal.