A study of a local-features-aware model for the problem of phase reconstruction from the magnitude spectrogram

This paper introduces a new algorithm for phase reconstruction from a modified magnitude spectrogram. Conventional methods are based on specific optimization process and classical overlap-add procedure. By referring to previous works, it is observed that use of local patterns may also provide good reconstructions. This leads to a reconstruction technique based on iterative inference of local patterns. The resulting method is suitable for incremental and real-time processing. It may be used directly, or as initial-guess to an optimization algorithm. The paper explores variants of models based local patterns. The results of these variants are compared.