Audio restoration based on multi-band spectral subtraction and missing data imputation

In this paper, an audio restoration method is proposed to reconstruct missing spectral components of noisy audio signals by a noise reduction method. In particular, the proposed method is based on multi-band spectral subtraction (MBSS) for noise reduction and missing data imputation (MDI) for audio restoration. In other words, noisy audio signals are first denoised by MBSS with weighted scale-factor bands. Then, distorted or missing audio spectral components are restored by MDI using a probabilistic latent component analysis. The performance of the proposed method is evaluated by using a log-spectral distortion (LSD) and compared with that of an audio denoising method based on block thresholding. It is shown from the comparison that the proposed method achieves lower LSD by 0.72 dB than the block thresholding method.