Abstract

This paper presents a novel phase reconstruction method (only from a givenamplitude spectrogram) by combining a signal-processing-based approach and adeep neural network (DNN). To retrieve a time-domain signal from its amplitudespectrogram, the corresponding phase is required. One of the popular phasereconstruction methods is the Griffin-Lim algorithm (GLA), which is based onthe redundancy of the short-time Fourier transform. However, GLA often involvesmany iterations and produces low-quality signals owing to the lack of priorknowledge of the target signal. In order to address these issues, in thisstudy, we propose an architecture which stacks a sub-block including twoGLA-inspired fixed layers and a DNN. The number of stacked sub-blocks isadjustable, and we can trade the performance and computational load based onrequirements of applications. The effectiveness of the proposed method isinvestigated by reconstructing phases from amplitude spectrograms of speeches.

 

Introduction (beta)

None

 

Conclusion (beta)

None