[SUGGESTION] ARDFTSRC - Real Discrete Fourier Transform Sample Rate Conversion

[MarcoRavich]

Hi there,
since this is the main .NET audio library on the scene, it would be great to add this very interesting resampling algorithm:

HA discussion

C++ port by @mrspoonsi: https://github.com/mrspoonsi/ARDFTSRC#readme

Quality tests by hydrogenaudio

• default
• minimum phase

Hope that inspires !

[mdcdi1315]

Sounds like a good idea to implement. I do not know about aliasing that will be caused by this, but I believe that the FFT transforms is using are relatively good.

An implementation problem is maybe how to implement an adequately fast and good FFT to use.

So +1 on this.

[MarcoRavich]

Well, according to the author (mycroft) in this post:

Note that even FFTW3 is not the fastest one here:

with float path processing: 44100 -> 48000, all single thread:

librempeg with ARDFTSRC filter + native TX ~ 1000x speed

librempeg with ARDFTSRC filter + fftw3 ~ 1700x speed

librempeg with ARDFTSRC filter + libhpk ~ 2500x speed (AVX2 only SIMD, there is also AVX512 if one wish to use it)

So single point of failure in both precision and speed/performance is how heavily optimized FFT/RDFT used library implementation is.
For pure N upsampling, and for relatively big N, where there is zero-padding of several factors this could also be optimized even more in theory.

fftw3 is https://www.fftw.org/
libhpk is https://hpkfft.com/

Hope that helps.

[MarcoRavich]

New Year bump.

ARDFTSRC reaches top scoring in official HA's SRC quality ranking !