Enhancing the EVS Codec in Wideband Mode by Blind Artificial Bandwidth Extension to Superwideband
Johannes Abel, Ernst Seidel, Tim Fingscheidt · 2018
In this paper we present a blind artificial bandwidth extension (ABE) framework for wideband-to-superwideband extension of speech signals coded by the recently standardized enhanced voice services (EVS) codec. The proposed solution is instrumentally evaluated using a deep neural network with and without a convolutional layer. Using a spectral distance metric, an improvement of more than 10 dB was obtained in the missing frequency band for the proposed ABE over the underlying wideband signal. A subjective comparison category rating (CCR) listening test between the underlying wideband speech and the generated superwideband speech revealed a highly significant improvement of 1.52 CMOS points.