DNN-based speech bandwidth expansion and its application to adding high-frequency missing features for automatic speech recognition of narrowband speech

We propose a number of enhancement techniques to improve speech quality in bandwidth expansion (BWE) from narrowband to wideband speech, addressing three issues, which could be critical in real-world applications, namely: (1) discontinuity between narrowband spectrum and the estimated high frequency spectrum, (2) energy mismatch between testing and training utterances, and (3) expanding bandwidth of out-ofdomain speech signals. With an inherent prediction of missing high frequency features in bandwidth-expanded speech we also explore the feasibility of adding these estimated features to those extracted from narrowband speech in order to improve the system performance for automatic speech recognition (ASR) of narrowband speech. Leveraging upon a recently-proposed deep neural network based speech BWE system intended for hearing quality enhancement these techniques not only improve over the traditionally-adopted objective and subjective measures but also reduce the word error rate (WER) from 8.67% when recognizing narrowband speech to 8.26% when recognizing bandwidthexpanded speech, and almost approaching the WER of 8.12% when recognizing wideband speech in the 20,000-word openvocabulary Wall Street Journal ASR task.