PaperlayerView exact revision on arXiv

Source-linked AI summary

A framework for invertible, real-time constant-Q transforms

Nicki Holighaus, Monika Dörfler, Gino Angelo Velasco, Thomas Grill

arXiv:1210.0084v1math.FA

TL;DR

Existing constant-Q transforms can be inefficient and lack invertibility, despite the need for non-linear frequency resolution in audio. The paper builds a frame-based, perfectly invertible constant-Q transform and processes slices for real-time use; it states that the sliced transform perfectly recovers the signal. The paper also notes that judging synthesis from modified coefficients is beyond its scope.

  • Problem

    Existing constant-Q transforms may be inefficient and non-invertible, although audio applications use non-linear frequency resolution.

    Source: §1 · p. 2

  • Method

    The framework uses nonstationary Gabor frames for a perfectly invertible constant-Q transform and slices the input for bounded-delay, linear-time processing.

    Source: §1 · p. 2

  • Results

    The paper states that applying its sliCQ synthesis to sliCQ coefficients perfectly recovers the input signal.

    Source: §4.2 · p. 9

  • Takeaways & Limitations

    The framework supports real-time invertible constant-Q processing and generalization to other nonlinear frequency scales such as mel or Bark.

    Source: §7 · p. 17

    The quality of synthesis from modified coefficients lacks a properly defined accuracy measure or target signal, and in-depth evaluation is beyond the article’s scope.

    Source: §6 · p. 17

Abstract

from arXiv · show

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 and in particular non-linear 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.

1. Introduction

The paper proposes a perfectly invertible constant-Q transform and a sliced version designed for real-time processing. It addresses the limitations of conventional time-frequency representations and existing constant-Q transforms.

1. Introduction

  • Constant-Q frequency spacing gives finer resolution at low frequencies and better time resolution at high frequencies, while STFT bins are uniformly spaced.The paper describes constant-Q resolution as better reflecting human auditory resolution and preserving overtone structures under frequency shifts.Source: §1 · p. 1
  • Classical CQTs may be computationally inefficient and lack invertibility, limiting unbiased reconstruction for extracting or modifying signal components.Earlier reconstruction methods are approximate.Source: §1 · p. 2
  • The CQ-NSGT uses nonstationary Gabor transforms to provide perfect invertibility, and slicing enables bounded-delay, linear-time sliCQ processing.The framework targets perfect reconstruction up to numerical precision.Source: §1 · p. 2
  • The CQ-NSGT requires a Fourier transform of the entire signal before processing; slicing addresses this, but coefficient modifications may produce blocking or time-aliasing.Slicing trades delay against finest frequency resolution, and its parameters can suppress artifacts while retaining the constant-Q coefficient structure.Source: §1 · p. 2

2. Nonstationary Gabor Frames

The section develops frame-based tools for stable reconstruction and adapts Gabor systems to nonuniform frequency resolution. Its frequency-adaptive painless NSG construction supports perfect reconstruction under stated frame conditions.

2. Nonstationary Gabor Frames

  • Frames generalize bases by allowing redundancy and design flexibility, and frame theory gives conditions for stable, perfect reconstruction.The discussion restricts the construction to finite, discrete signals, while noting that it extends to L2(R).Source: §2 · pp. 3–4

2.1. Frames.

  • Classical Gabor frames use one window shifted on a regular time-frequency lattice; relaxing that regularity yields nonstationary frames for adaptive resolution.The paper considers finite discrete signals and defines Gabor atoms through time and frequency shifts of a single window.Source: §2.1 · pp. 4–5, §2 · p. 4

2.2. Frequency-Adaptive Painless Nonstationary Gabor Frames.

  • The method adapts bandwidth and sampling density in frequency by applying a nonstationary Gabor system to the input signal’s Fourier transform.Its frequency-domain filters may be centered at irregularly or geometrically spaced frequencies.Source: §2.2 · p. 5
  • Under the painless-frame conditions, dual NSG frames provide perfect reconstruction, and the canonical dual is available for the constructed system.The frame construction requires sufficiently dense time sampling; its coefficients are computed from frequency-domain windows.Source: §2.2 · pp. 5–6

3. The CQ-NSGT Parameters: Windows and Lattices

The CQ-NSGT uses frequency-adaptive filters to obtain constant-Q resolution, with center frequencies and bandwidths set from frequency and sampling parameters. Its lattice must meet a support condition for perfect reconstruction.

3. The CQ-NSGT Parameters: Windows and Lattices

  • For bins in a frequency range, constant-Q filters have approximately the same ratio of center frequency to support length.The filters are centered at their corresponding frequencies.Source: §3 · p. 7
  • Minimum and maximum frequencies, sampling rate, and bins per octave determine the center-frequency design, with the highest center below Nyquist.Source: §3 · p. 7
  • Interior bandwidths use neighboring center frequencies to set a constant Q-factor; endpoint bandwidths use their center frequencies and Q.Symmetric filters for real-valued signals and filters at zero and Nyquist ensure coverage of the frequency axis.Source: §3 · p. 7
  • Each interior filter is a sampled, translated, and dilated version of H, while plateau functions handle zero and Nyquist.Source: §3 · p. 7
  • The lattice supports perfect reconstruction when L/a_k is integer-valued and at least the support length L_k for every filter.Synthesis can use a dual frame, including the canonical dual.Source: §3 · p. 7
  • The chosen zero-padding for the slicing windows produces half-overlap.Source: Figure 2 · p. 8

4. Real-time processing and the sliCQ

The sliCQ approach applies the CQ-NSGT to overlapping signal slices to support real-time processing while preserving perfect reconstruction. Its coefficients approximate those of the full-length CQ-NSGT under appropriate slicing-window choices.

4.1. Structure of the sliCQ transform.

  • Overlapping length-2N slices are transformed separately, their coefficients arranged in a two-layer array, and synthesis uses dual-frame inversion followed by windowed overlap-add.Source: §4.1 · p. 8

4.2. Computation of a sliced constant-Q NSGT.

  • sliCQ synthesis perfectly recovers the input from its coefficients using the specified synthesis algorithm.Source: §4.2 · p. 9

4.3. The relation between CQ-NSGT and sliCQ.

  • A uniform partition-of-unity slicing window favors approximation of full-length CQ coefficients; the implementation uses zero-padded Tukey windows with essential length N and transition length M.Source: §4.3 · p. 10
  • More precise approximation-error estimates are beyond the paper's scope, so approximation quality is evaluated numerically in Section 5.3.Source: §4.3 · p. 11
  • The sliCQ spectrogram is defined as the squared magnitude of the sum of its two layers, whose coefficients share time-frequency positions and are processed together.Source: §4.3 · p. 11

5. Numerical Analysis and Simulations

The section analyzes computation cost, runtime, and how closely sliCQ coefficients approximate full-length CQ-NSGT coefficients. It reports linear-cost sliCQ processing and approximation behavior across signal sets and parameter choices.

5.1. Computation Time and Computational Complexity.

  • CQ-NSGT analysis and synthesis have O(L log L) complexity, while sliCQ has O(L) complexity when slice length is independent of signal length.The CQ-NSGT comparison with classical constant-Q implementations is referenced to [21].Source: §5.1 · pp. 11–12

5.2. Performance evaluation.

  • sliCQ computation time scales linearly with signal length, with deviations associated with unfavorable FFT lengths and performance varying with slice length.The authors attribute improvements at longer slice lengths to MATLAB’s more advanced internal FFT algorithm compared with the current sliCQ implementation.Source: §5.2 · p. 12, Figure 5 · p. 13

5.3. Approximation properties.

  • Minimal filter lengths below 8 samples produce visibly different representations, whereas lengths above 16 samples yield coefficients largely equivalent to a constant-Q transform.The approximation was evaluated on random complex signals and 90 music samples spanning pop, rock, jazz, and classical genres.Source: §5.3 · p. 13
  • Slice length has little influence on approximation, while transition-area ratio matters chiefly for moderately localized filters and results are largely independent of signal class.The authors relate transition-area effects to circular overspill; Set 2 outliers are attributed to signals especially sparse or dense in low-frequency regions.Source: §5.3 · pp. 13–14

6. Experiments on Applications

The application experiments use CQ-NSGT coefficient masks to isolate signal components and transpose a selected component. The authors present the synthesized results as illustrative, while noting that rigorous evaluation of modified-coefficient synthesis is beyond the paper’s scope.

6. Experiments on Applications

  • Masking CQ-NSGT coefficients can extract or suppress components, and transposition of a harmonic structure corresponds to shifting its spectrum along frequency bins.The experiment applies these procedures to modify a portion of a Glockenspiel signal.Source: §6 · pp. 14–15
  • The experiment masks transient, sinusoidal, and remainder portions, shifts the sinusoidal component upward by 2 semitones (8 frequency bins), then synthesizes the combined signal.The same mask is used for CQ-NSGT and sliCQ transpositions, with sliCQ examples at different slice lengths.Source: §6 · pp. 16–17
  • The examples indicate synthesis in accordance with intuition, but evaluating modified-coefficient synthesis lacks a properly defined accuracy notion or target signal and is beyond the article’s scope.The stated evaluation difficulty concerns modified coefficients rather than mere reconstruction.Source: §6 · p. 17

7. Summary and Conclusion

The paper concludes that its frame-theoretic framework supports real-time implementation of an invertible constant-Q transform and can generalize to other nonlinear frequency scales. It identifies coefficient-processing phase effects and time-frequency adaptivity as subjects for future work.

7. Summary and Conclusion

  • The framework implements a real-time invertible constant-Q transform using frame theory and generalizes to nonlinear frequency scales such as mel and Bark.The conclusion also reports numerical evaluation of computation time and approximation quality of true NSGT coefficients.Source: §7 · p. 17
  • Undesired phasing effects can arise when a signal component contributes to several adjacent filters, and the authors leave detailed investigation for future work.They also identify efficient adaptivity in both time and frequency, via varying preprocessing-window lengths, as future work.Source: §7 · p. 17

8. Appendix

The appendix derives properties of CQ-NSGT and sliCQ. Its proof passages connect the frame assumptions and synthesis steps to reconstruction, and relate slice-wise coefficients to full-transform coefficients.

8.1. Derivation of CQ-NSGT properties.

  • The CQ-NSGT proof uses bounded frame sums and the stated filter-support condition to construct the inverse frame operator and complete the reconstruction argument.The appendix then invokes dual NSG frames and an inverse discrete Fourier transform in the synthesis proof.Source: §8.1 · pp. 18–19

8.2. Derivation of sliCQ properties.

  • The sliCQ proofs express reconstructed slice outputs through the CQ-NSGT coefficients and derive coefficient relations from sampling and periodization of the filters.One proof bounds the coefficient difference using Cauchy–Schwarz; another identifies the inverse Fourier transform of sampled filters with a periodization.Source: §8.2 · pp. 19–20
Loading 1210.0084v1…