Audio super-resolution using analysis dictionary learning

Super-resolution is an important problem in signal processing. It aims to reconstruct a high-resolution (HR) signal from a low-resolution (LR) input. We consider the super-resolution problem for audio signals in the time-frequency domain and propose a method using analysis dictionary learning. The input to our proposed method is the LR spectrogram matrix of an audio signal, where some rows corresponding to high-frequency information are lost. First, an analysis dictionary is learned from the spectrogram of some related audio signals. The learned dictionary is then applied in an ℓ1-norm regularization term for the reconstruction of the HR spectrogram. Experimental results with piano signals demonstrate the advantage of the learned dictionaries in reconstructing HR spectrograms.