17 April 2015 in IEEE Transactions on Signal Processing
We consider the problem of designing spectral graph filters for the construction of dictionaries of atoms that can be used to efficiently represent signals residing on weighted graphs. While the filters used in previous spectral graph wavelet constructions are only adapted to the length of the spectrum, the filters proposed in this paper are adapted to the distribution of graph Laplacian eigenvalues, and therefore lead to atoms with better discriminatory power. Our approach is to first characterize a family of systems of uniformly translated kernels in the graph spectral domain that give rise to tight frames of atoms generated via generalized translation on the graph. We then warp the uniform translates with a function that approximates the cumulative spectral density function of the graph Laplacian eigenvalues. We use this approach to construct computationally efficient, spectrum-adapted, tight vertex-frequency and graph wavelet frames. We give numerous examples of the resulting spectrum-adapted graph filters, and also present an illustrative example of vertex-frequency analysis using the proposed construction.
1 October 2011 in Journal of Computational and Applied Mathematics
Signal analysis with classical Gabor frames leads to a fixed time–frequency resolution over the whole time–frequency plane. To overcome the limitations imposed by this rigidity, we propose an extension of Gabor theory that leads to the construction of frames with time–frequency resolution changing over time or frequency. We describe the construction of the resulting nonstationary Gabor frames and give the explicit formula for the canonical dual frame for a particular case, the painless case. We show that wavelet transforms, constant-Q transforms and more general filter banks may be modeled in the framework of nonstationary Gabor frames. Further, we present the results in the finite-dimensional case, which provides a method for implementing the above-mentioned transforms with perfect reconstruction. Finally, we elaborate on two applications of nonstationary Gabor frames in audio signal processing, namely a method for automatic adaptation to transients and an algorithm for an invertible constant-Q transform.
20 December 2012 in IEEE Transactions on Audio, Speech, and Language Processing
Audio signal processing frequently requires time-frequency representations and in many applications, a non-linear spacing of frequency bands is preferable. This paper introduces a framework for efficient implementation of invertible signal transforms allowing for non-uniform frequency resolution. Non-uniformity in frequency is realized by applying nonstationary Gabor frames with adaptivity in the frequency domain. The realization of a perfectly invertible constant-Q transform is described in detail. To achieve real-time processing, independent of signal length, slice-wise processing of the full input signal is proposed and referred to as sliCQ transform. By applying frame theory and FFT-based processing, the presented approach overcomes computational inefficiency and lack of invertibility of classical constant-Q transform implementations. Numerical simulations evaluate the efficiency of the proposed algorithm and the method's applicability is illustrated by experiments on real-life audio signals .
14 October 2019 in IEEE/ACM Transactions on Audio, Speech, and Language Processing
We study the ability of deep neural networks (DNNs) to restore missing audio content based on its context, i.e., inpaint audio gaps. We focus on a condition which has not received much attention yet: gaps in the range of tens of milliseconds. We propose a DNN structure that is provided with the signal surrounding the gap in the form of time-frequency (TF) coefficients. Two DNNs with either complex-valued TF coefficient output or magnitude TF coefficient output were studied by separately training them on inpainting two types of audio signals (music and musical instruments) having 64-ms long gaps. The magnitude DNN outperformed the complex-valued DNN in terms of signal-to-noise ratios and objective difference grades. Although, for instruments, a reference inpainting obtained through linear predictive coding performed better in both metrics, it performed worse than the magnitude DNN for music. This demonstrates the potential of the magnitude DNN, in particular for inpainting signals that are more complex than single instrument sounds.
27 February 2018 in IEEE/ACM Transactions on Audio, Speech, and Language Processing
We present a novel method for the compensation of long duration data loss in audio signals, in particular music. The concealment of such signal defects is based on a graph that encodes signal structure in terms of time-persistent spectral similarity. A suitable candidate segment for the substitution of the lost content is proposed by an intuitive optimization scheme and smoothly inserted into the gap, i.e., the lost or distorted signal region. Extensive listening tests show that the proposed algorithm provides highly promising results when applied to a variety of real-world music signals.