Want programmer for resampling experiment...

2004-05-09 07:50:07

Hi,

My understanding of one of the 'audiophile' claims made re: upsampling is that it supposedly can reduce HF ringing by altering the filter cutoff slope (or moving it to a different frequency, outside audible range).

I'm curious to test an SSRC-based plugin for either WinAMP 2/5 or Foobar, with these alterations:

* Selection of (A) normal, steep brickwall reconstruction filter for rejecting sampling images over fs/2, (B) gradual slope reconstruction filter for image rejection, and (C ) no reconstruction filter, leaving the ultrasonic images (when upsampling) alone.

I want to see what kind of subjective differences (if any) can be heard with my M-Audio AP 24/96 soundcard when upsampling from 44.1/16 to 88.2/24 or 96/24 using different sorts of HF image filters, including none.  Technically this would be faulty resampling (as I understand it) but this is a matter of curiosity.

If any programmers are interested enough to whip something up based on already-existing code (and maybe participate in a listening test) please contact me or post to the thread.  Preferable would be a realtime output plugin, but just a change to the SSRC command line code would be sufficient for testing (unfortunately, I don't remember enough 'C' even to do this much myself).  If this is a dumb idea due entirely to my lack of understanding of resampling, I apologize ... thanks.

Want programmer for resampling experiment...

Reply #1 – 2004-05-09 16:12:42

The amount of aliasing present without any filter will directly depend on the process used to upsample. Since you start from 44.1 kHz, there is no content at all above 22050 Hz. What will appear above 22050 Hz will only be generated by an imprefect upsampling process : many artifacts if you use nearest neighbour, near perfect lowpass if you use high order polynomial interpolation... Ideally, you'll get the exact filter that was used to record the original 44.1 kHz file.

You can look at the three first pictures in paragraph 3 on this page : http://perso.numericable.fr/~laguill2/cdr/cdr.htm

They are three upsampled versions (44.1 -> 96 kHz ) of the first spectrogram in the page. All were done without any antialias filter, but with different qualities in SoundForge 4.5. We can see that with the lowest quality, the low frequency content is mirrored above 22050 Hz, while with the highest quality, the original frequency content (0-22.05 kHz) is kept unchanged.

Want programmer for resampling experiment...

Reply #2 – 2004-05-10 12:29:05

That's true if you generate the extra samples using some specific algorithm.

However, in "classical" upsampling (which is what fewtch seems to be thinking of) you only consider integer resampling ratios, and simply zero pad as necesary between the original samples (resulting in image frequencies above the original fs/2) and then filter.

You can't do this from 44.1kHz to 96kHz.

Cheers,
David.