| Based on conventional audio bandwidth extension (BWE) techniques, the spectrum correlation between high frequency (HF) and low frequency (LF) was analyzed in this thesis. First, the effective estimation method of HF subband energy was proposed according to the spectral envelop and the timbre features of LF components. Second, the nonlinear method about fine structure of HF was presented based on phase space reconstruction from nonlinear dynamics and nearest neighbour matching. Finally, a blind BWE of audio signals from wideband to super wideband audio is implemented in a real codec.In order to estimate the HF subband energy, the Gaussian mixture model (GMM) is introduced for modeling the joint probability density of HF and LF features. The GMM-based HF subband energy estimation method was proposed using minimum mean square error (MMSE) criterion, and the estimation accuracy is guaranteed for stationary audio data. Furthermore, the hidden Markov model (HMM) is introduced into the aforementioned BWE method for improving GMM-based HF subband energy estimation method. The perceived quality of extended audio is improved, and the burst of HF energy are eliminated for the complicated audio data.In order to recover the fine structure of HF, the phase space reconstruction is employed to verify the nonlinear characteristic of audio signal in frequency domain. According to local evolving tendency of LF trajectory, a nonlinear extension of the fine structure in HF band was proposed based on nearest neighbour matching method.In order to evaluate the performance of the proposed methods, a blind audio bandwidth extending algorithm from wideband to super wideband audio is build by combining HMM with nonlinear prediction, and it has been applied into G722.1 wideband audio codec at the bit rate of 24kbit/s. The objective and subjective quality tests show that the performance of G.722.1 with the proposed algorithm is comparable with G.722.1C super wideband codec at the same bit rate. |
PDF Full Text Request