Investigation of DNN Prediction of Power Spectral Envelopes for Speech Coding & ASR
Christine Pickersgill, Stephen So, Belinda Schwerin · 2018
This paper proposes a DNN-based preprocessing method for speech coding and automatic speech recognition applications. The method proposed here maps noisy log power spectra to “clean” smoothed log power spectral envelopes using DNN pre-diction. The proposed method has the advantage of combining feature extraction with DNN-based enhancement, thus reducing computational time and resources. The TIMIT speech database with various additive noise types was used to train the DNN, and the NN prediction results are compared to the target clean log power spectral envelopes using log spectral distortion. The proposed method is found to have lower log spectral distortion measurements compared to similar neural networks that map noisy power spectra to clean power spectra.