Artificial bandwidth extension using deep neural networks for spectral envelope estimation

Many artificial speech bandwidth extension (ABE) approaches perform source-filter decomposition of the input narrowband speech, with subsequent computation of upper frequency band (UB) spectral envelope posteriors. In this paper we perform a direct comparison of HMM- and deep neural network (DNN)-based modeling of likelihoods or posteriors for ABE UB envelope estimation. DNN-based approaches turn out to significantly exceed GMM-based ones in speech quality. Further analysis reveals that this is not due to a better shape of the estimated UB spectral envelope, but primarily due to a much better estimate of the energy ratio of the upper band vs. the lower band - an important result with significant impact on ABE speech quality particularly for fricative sounds.