부산대학교 도서관

This paper proposes a dual-branch method to speech bandwidth extension in spectral domain. The network architecture proposed in this study includes a magnitude branch and a complex branch. The former is intended to provide an approximate estimation of the magnitude of high-frequency speech signals, while the latter is designed in parallel and capable of estimating phase information as well as residual magnitudes, which can account for compensation of the magnitude branch. To capture long-term time-frequency dependencies more effectively, both branches of our network employ convolutional recurrent networks(CRN). The proposed method outperforms the baselines in terms of Log-Spectral Distance (LSD), Perceptual Evaluation of Speech Quality (PESQ), and Scale-Invariant Signal-to-Noise Ratio (SI-SNR).