Wide-band speech coding using kernel methods and bandwidth extension based on parametric stereo

摘要

A novel wide-band speech coding scheme is developed, in this paper, based on kernel methods and bandwidth extension. The KLMS algorithm, a kernelized version of the well-known LMS algorithm, is employed in the framework of the backward ADPCM technique for encoding the narrow-band part of the wide-band speech. Simulation results show that utilizing this nonlinear method results in an average improvement of up to 3.4 dB in the SNR and 0.28 in the PESQ measure of the decoded speech. The resultant narrow-band codec is subsequently extended to the wide-band speech using a novel bandwidth extension technique inspired by the parametric stereo coding. It is shown that the KLMS algorithm is also effective in this framework. This leads to a wide-band speech coding scheme built on the nonlinear narrow-band codec at the extra cost of a small increase in the bite rate.