A novel method of artificial bandwidth extension using deep architecture

This paper presents a novel artificial bandwidth extension (ABE) framework based on deep neural networks (DNNs) with a multiple-layer’s deep architecture. It demonstrates the suitability of DNNs for modeling log power spectra of speech signals using the application of ABE. The DNN is used to estimate the log power spectra in the high-band. Two strategies are proposed to improve the performances of the proposed ABE system. First, global variance equalization is proposed to alleviate the over-smoothing issue in generated log spectra. Second, rich acoustic features in the low-band are considered to improve the construction of the log power spectra in the high-band. Experimental results demonstrate that the proposed framework can achieve significant improvements in both objective and subjective measures over the different baseline methods.