Phase reconstruction of spectrograms with linear unwrapping: application\n to audio signal restoration

2016/05/24 by Paul Magron, Magron, Paul, Roland Badeau +3
Computer Science · #Image and Signal Denoising Methods #Blind Source Separation Techniques #Speech and Audio Processing

paper · pdf · doi:10.48550/arxiv.1605.07467

Abstract

This paper introduces a novel technique for reconstructing the phase of\nmodified spectrograms of audio signals. From the analysis of mixtures of\nsinusoids we obtain relationships between phases of successive time frames in\nthe Time-Frequency (TF) domain. To obtain similar relationships over\nfrequencies, in particular within onset frames, we study an impulse model.\nInstantaneous frequencies and attack times are estimated locally to encompass\nthe class of non-stationary signals such as vibratos. These techniques ensure\nboth the vertical coherence of partials (over frequencies) and the horizontal\ncoherence (over time). The method is tested on a variety of data and\ndemonstrates better performance than traditional consistency-based approaches.\nWe also introduce an audio restoration framework and observe that our technique\noutperforms traditional methods.\n

Related

• Signal Reconstruction from Mel-spectrogram Based on Bi-level Consistency of Full-band Magnitude and Phase
• Model-based STFT phase recovery for audio source separation
• An Explicit Consistency-Preserving Loss Function for Phase Reconstruction and Speech Enhancement
• Phase reconstruction based on recurrent phase unwrapping with deep neural networks
• Generative adversarial network-based approach to signal reconstruction from magnitude spectrograms