Audio Bandwidth Extension Based On Grey Model And Support Vector Machine

On the basis of traditional audio bandwidth extension methods and nonlinearanalysis and prediction of audio signals, this thesis studies the correlation between thehigh and low frequencies of super wideband audio signals according to the evolvingtrends of spectral envelope, and proposes a kind of high-frequency sub-band energyestimation method based on grey model. In addition, in the light of the nonlinearity ofaudio spectrum, the trajectory of the high-frequency phase points is nonlinearlypredicted by using phase space reconstruction, and least square support vectormachine to effectively restore the high-frequency fine spectrum. Finally, the wholeblind bandwidth extension method is applied into an actual audio codecs to implementthe bandwidth extension from wideband to super wideband audio.For the estimation of high-frequency sub-band energy, this thesis firstly studiesthe applicability of grey model on the estimation of high-frequency sub-band energy.According to the correlation and evolving trends between the high and lowfrequencies of audio signals, the accumulated generating and background valuegenerating operations are adopted to reduce the randomness of the low-freuqencyspectral envelope series. Secondly, the GM(1,1) prediction model is build up by usingleast square fitting to effectively estimate the high-frequency sub-band energy. Inaddition, the generating methods of background value for grey model are also studiedfor further optimalizing the prediction model.Furthermore, a grey Verhulst model is presented to estimate the high-frequencysub-band energy to describe the approximately saturated characteristics of the shapeof spectral envelope in terms of grey Verhulst differential equation. According to theperformance of two proposed grey model for extending the bandwidth of audiosignals, an optional prediction model is proposed to adaptively estimate thehigh-frequency sub-band energy under the minimum fitting error criterion, and theperformance of spectral envelop extension is effectively improved.For the estimation of high-frequency fine spectrum, a delay reconstructionmethod is employed to convert the one-dimensional spectral series into themulti-dimensional phase space, according to the nonlinearity of audio spectrum, andlocal least square support vector machine is introduced to nonlinearly predict thehigh-frequency fine spectrum. Combining with optional prediction based on greymodel, the whole blind bandwidth extension method is realized from wideband tosuper wideband for audio signals.In order to verify the extension performance, the proposed bandwidth extensionmethod is applied into the wideband audio codec of G.722.1at the bitrate of24kbps,and its performance is evaluated in comparison with the reference method that is based on Gaussian mixture model and nearest neighbor mapping and the superwideband audio codec G.722.1C. Test results show that the super wideband audiosignals reconstructed by the proposed method retain the most spectral characteristicsof original audio signals. The boundary between the high and low frequencies ofaudio spectrum is naturally transited, and the auditory perception is smooth. Theperformance of subjective and objective tests shows that the proposed bandwidthextension method is comparable to G.722.1C codec and gains remarkablyimprovement on the reference method.