Bandwidth extension of audio signal based on SOM prediction model

In this paper, a blind bandwidth extension approach is proposed based on prediction model of self-organizing map (SOM). Owing to nonlinear characteristics of audio spectrum series, the fine spectrum of low-frequency (LF) information is first described based on the principles of nonlinear dynamic, and then phase points are derived from reconstructing fine spectrum in phase space. Next, these phase points are trained by SOM algorithm to obtain the SOM structure. Thus, the fine structure of high-frequency spectrum can be recovered with SOM prediction model. Combining with Gaussian mixture model (GMM) to adjust spectral envelope of extended high-frequency (HF) components, the bandwidth of wideband audio signal is extended to super-wideband. Both subjective and objective testing results demonstrate that the proposed method outperforms the reference approaches in most cases.