[Paper Review] A Reduced Multiple Gabor Frame for Local Time Adaptation of the Spectrogram

Marco Liuni, Axel Röebel+2 more|arXiv (Cornell University)|Sep 27, 2011

Image and Signal Denoising Methods8 references4 citations

TL;DR

This paper proposes a reduced multiple Gabor frame for local time-adaptive spectrogram analysis, using Rényi entropy to select optimal window sizes per time segment. By minimizing entropy across a set of scaled Gabor frames, the method achieves variable-resolution time-frequency representations with perfect reconstruction via a dual-frame-based phase vocoder, improving on fixed-resolution and non-reconstructible adaptive methods.

ABSTRACT

In this paper we propose a method for automatic local time adap- tation of the spectrogram of an audio signal, based on its decomposition within a Gabor multi-frame. The sparsity of the analyses within each individual frame is evaluated through the Rényi entropies measures. According to the sparsity of the decompositions, an optimal resolution and a reduced multi-frame are determined, defining an adapted spectrogram with variable resolution and hop size. The composition of such a reduced multi-frame allows an immediate definition of a dual frame: re-synthesis techniques for this adapted analysis are easily derived by the traditional phase vocoder scheme.

Motivation & Objective

• To address the limitation of fixed-resolution spectrograms in capturing both transient and tonal components in audio signals.
• To overcome the need for manual window selection by enabling automatic, signal-adaptive resolution adjustment.
• To maintain perfect signal reconstruction while reducing redundancy in multiple Gabor frame systems.
• To develop a computationally efficient framework that allows local time adaptation without sacrificing reconstruction fidelity.

Proposed method

• The method constructs a multiple Gabor frame using a finite set of scaled window functions $ g^l(t) = \frac{1}{\sqrt{l}} g\left(\frac{t}{l}\right) $ for $ l \in L \subset \mathbb{R}^+ $, enabling variable time-frequency resolution.
• For each analysis segment, Rényi entropy of order $ \alpha = 0.7 $ is computed over the spectrogram to evaluate sparsity, with lower entropy indicating better concentration and thus optimal resolution.
• The window size with minimal Rényi entropy is selected per segment, forming a reduced multi-frame that preserves the redundancy and overlap structure of the original frames.
• A dual frame is implicitly defined through the phase vocoder re-synthesis technique, enabling perfect reconstruction via weighted summation of overlapping analysis segments.
• The method allows variable hop sizes per segment by selecting only the optimal frame per segment, avoiding unnecessary short hops.
• Re-synthesis is achieved using the least-squares error estimation framework from [8], extended to variable windows, ensuring signal recovery with minimal error.

Experimental results

Research questions

• RQ1 How can time-frequency resolution be adaptively adjusted per time segment to better represent transient and tonal components in audio signals?
• RQ2 What sparsity measure can effectively identify the optimal window size for local signal features in a multi-resolution Gabor frame?
• RQ3 Can a reduced multi-frame be constructed from multiple Gabor frames that maintains perfect reconstruction capability?
• RQ4 How can variable hop sizes be incorporated into adaptive spectrogram analysis without increasing computational cost?
• RQ5 What is the impact of entropy minimization on the perceptual quality and accuracy of the resulting spectrogram?

Key findings

• The algorithm successfully adapts spectrogram resolution in real-time, selecting the optimal window size per segment based on Rényi entropy minimization.
• For a marimba B4 note, the method achieves high time resolution at attack and high frequency resolution during sustain, outperforming fixed-window STFTs.
• The pre-echo artifact present in the 4096-sample Hanning window spectrogram is completely removed in the adaptive spectrogram.
• The synthetic frequency-modulated sinusoid example confirms that small windows are selected in regions of rapid frequency change, while large windows are used in stationary regions.
• Perfect reconstruction is achieved using the phase vocoder re-synthesis method with the reduced multi-frame, ensuring signal fidelity.
• The method allows user-defined segment duration and overlap, enabling control over the smoothness of resolution transitions.