Development of the CUHK Dysarthric Speech Recognition System for the UASpeech Corpus

Dysarthric speech recognition is a highly challenging task. The articulatory motor control problems associated with neuromotor conditions produce large mismatch against normal speech, In addition, such data is difficult to collect in large quantities. This paper presents the development of the Chinese University of Hong Kong automatic speech recognition (ASR) system for the Universal Access Speech (UASpeech) [1]. A range of deep neural network (DNN) acoustic models and their more advanced variants based on time delayed neural networks (TDNNs) and long short-term memory recurrent neural networks (LSTM-RNNs) were developed. Speaker adaptation by learning hidden unit contributions (LHUC) was used. A semi-supervised complementary auto-encoder system was further constructed to improve the bottleneck feature extraction. Two out-of-domain (OOD) ASR systems separately trained on broadcast news and switchboard data were cross domain adapted towards the UASpeech data and adopted in system combination. The final combined system gave an overall word accuracy of 69.4% on the 16-speaker test set.