Deep Neural Network based Complex Spectrogram Reconstruction for Speech Bandwidth Expansion
Hongjiang Yu, Wei‐Ping Zhu · 2020
In this paper, we present a deep neural network (DNN) based complex spectrogram reconstruction algorithm for speech bandwidth expansion, where the DNN is applied for estimating the real and imaginary parts of spectrograms of the wideband speech from those of the narrowband speech. Unlike the previous DNN based method, which only estimates the magnitude and employs the simple mirror version phase for reconstruction, we employ the complex spectrogram to recover the magnitude and phase of the high-frequency component simultaneously. Experimental results demonstrate that our proposed method outperforms the non-negative matrix factorization (NMF) and the state-of-the-art DNN based speech bandwidth expansion methods in terms of objective performance metrics.