Non-linear harmonic generation based blind bandwidth extension considering aliasing artifacts

Abstract

This paper has three aims that are to point out that signals generated by conventional BWE methods generally include some aliasing artifacts, to propose a novel bandwidth extension (BWE) method considering the effects of aliasing artifacts, and to apply various BWE methods to speaker verification to evaluate the effectiveness of the BWE ones. Study on BWE methods using non-linear functions has a long history started from for analog signal processing, but conventional BWE ones have never considered the influence of aliasing artifacts caused by the band limitation that digital signals have. This paper is among the first to point out that discrete-time signals generated by BWE methods generally include some aliasing artifacts due to the band limitation to be decided according to the sampling frequency. Next, a new non-linear artificial BWE method, considering of aliasing artifacts, is proposed. Moreover, to evaluate the proposed framework, speaker verification experiments and objective tests are conducted. Experiment results show that speech signals extended by the proposed framework provide the error reduction of 46.2%, compared with a typical conventional method. Additionally, the generated speech signals are evaluated by using three object measures: PESQ, RMS-LSD and STOI. It is also implied that equal error rates in speaker verification tasks have a closest relation with RMS-LSD in with another measures.