ConstantQ Transform?

[constd]

My search in the algorithm reference documentation
and a quick search of the repository proved fruitless.
Is there an implementation of it (e.g. like this) available in Essentia?

[dbogdanov]

We do not have this transform, although it could be nice to have it if we see benefits of using it instead of FFT for descriptor computation.

[dbogdanov]

Approach by Schörkhuber and Klapuri 2010:
https://code.soundsoftware.ac.uk/projects/constant-q-cpp

[nkundiushuti]

This is an invertible constant q based on gabor frames, implemented in python and matlab:
https://github.com/alexbw/nsgt
http://www.univie.ac.at/nonstatgab/cqt/

[dbogdanov]

C++ implementation in QM DSP Library

[dbogdanov]

Implemented in the latest master.

[dbogdanov]

We are in the process of testing our current implementation.

Alternative implementations:
https://github.com/librosa/librosa/blob/master/librosa/core/constantq.py
http://librosa.github.io/librosa/generated/librosa.core.cqt.html
http://librosa.github.io/librosa/_modules/librosa/core/constantq.html#cqt

[jordipons]

Summary of the most relevant CQT implementations:

• Brown, 1991, first proposed CQT with a naive very slow implementation.
• Brown and Puckette, 1992: implemented a faster CQT based on a sparse representation in the frequency domain. This is the current Essentia implementation!
• Schörkhuber and Klapuri, 2010: a faster CQT based on the same principles as Brown and Puckette, 1992. For first time, it is introduced an algorithm for an approximated reconstruction of the CQT coefficients. Code: http://www.iem.at/~schoerkhuber/cqt2010/
• Velasco, Holighaus, Dörfler and Grill, 2011: they approach the problem differently - by means of a nonstationary Gabor transform. This allows perfect reconstruction for first time while the transform is still computationally efficient (faster than Schörkhuber and Klapuri, 2010). However, it does not allow real-time implementations and phases are not accurate. Code: http://www.univie.ac.at/nonstatgab/toolbox.php
• Holighaus, Dörfler, Velasco and Grill, 2012: Based on the Velasco, Holighaus, Dörfler and Grill, 2011 - allowing perfect reconstruction. They propose sliCQT (slicing by using an overlapping window) to allow real-time computations. Code: http://www.univie.ac.at/nonstatgab/toolbox.php
• Schörkhuber, Klapuri, Holighaus and Dörfler, 2014: Based on the Velasco, Holighaus, Dörfler and Grill, 2011 - allowing perfect reconstruction. They solve the problem with phases by means of a frequency mapping and they also propose a Variable-Q transform (that allows ie. ERBlets). Code: http://www.cs.tut.fi/sgn/arg/CQT/

Schörkhuber, Klapuri, Holighaus and Dörfler, 2014, a nonstationary Gabor transform, allows perfect reconstruction while the phases are still accurate. It might be interesting to implement this in Essentia.

[dbogdanov]

@pabloEntropia we have to select among the last two approaches (Holighaus 2012 and Schörkhuber 2014) or their combination. You can read these papers meanwhile.

[palonso]

Great I will start reading them!

[dbogdanov]

@pabloEntropia what's your feedback on these two papers?

[palonso]

Both papers are based on CQ-NSGT (Velasco, Holighaus, Dörfler and Grill, 2011). Holighaus, Dörfler, Velasco and Grill, 2012 just shows how to use the CQ-NSGT frame-wise (sliCQ).
I think that the interesting algorithm to implement is Schörkhuber, Klapuri, Holighaus and Dörfler, 2014, for the reasons exposed by @jordipons
After doing this, we should decide if we create a second algorithm relaying on our CQ-NSGT implementation, Windowing and FrameGenerator to achieve the sliCQ frame-wise behavior or just create a python example script explaining it.

[asapsmc]

Hi!
Any plans for the Invertible-CQT?
Thanks.

[constd]

just saw this: https://mtg.github.io/essentia-labs/news/2019/02/07/invertible-constant-q/
which is awesome! Thank you for taking the time to implement the an invertible CQT!

[mauriciovmc]

Hello!

Planning to use the variable-Q transform here with Python routines. Is there an implementation for Python of it available? Otherwise, I'll just use the Matlab version instead...

Thanks :)

[palonso]

In the aforementioned post we show how to use our Constant-Q C++ implementation with Python wrappers in just a few lines:)

[sevagh]

@palonso

Thanks for this work - can you confirm that the "Schörkhuber 2014" (i.e. the "best" CQ-NSGT with phase) is the implementation in the code? Also, did you have a chance to implement the time-aligned coefficients/matrix form (namely the rasterization/interpolation as described by the paper?)

In the code it looks like this is true:

• essentia/src/algorithms/standard/nsgconstantq.cpp

  Line 37 in 07d089f

  " [1] Schörkhuber, C., Klapuri, A., Holighaus, N., & Dörfler, M. "

• essentia/src/algorithms/standard/nsgiconstantq.cpp

  Line 37 in 07d089f

  " [1] Schörkhuber, C., Klapuri, A., Holighaus, N., & Dörfler, M. "

• essentia/src/algorithms/standard/nsgconstantqstreaming.cpp

  Line 35 in 07d089f

  " [1] Schörkhuber, C., Klapuri, A., Holighaus, N., & Dörfler, M. (n.d.). A Matlab Toolbox for Efficient Perfect Reconstruction Time-Frequency Transforms with Log-Frequency Resolution.");

However in the associated feature release/blog post (https://mtg.github.io/essentia-labs/news/2019/02/07/invertible-constant-q/), the Schörkhuber 2014 paper is not mentioned in the references.