I was looking for a 32KHz VS 44KHz listening test performed on real music samples, but had no luck. I only found some how-high-can-i-hear tests performed on generated sine waves.
My point is that even if someone is able to "hear" (IMO it's rather feeling, or sensing) synthetic sine waves of 16KHz or higher frequencies, he/she probably wouldn't hear anything that high in real music. Not that there are no such a high frequencies in real music, but I think that we are not sensitive enough to spot them in case when they are covered by frequencies in a more human range.
It would be useful to know that, for example in lossy encodings I could resample 44KHz to 32KHz, so no bits would be wasted on encoding something I won't hear anyway.
The way the test could be performed:
1. Pick some song and rip it to lossless file A
2. View its spectrogram and make sure that it contains strong signals above 16KHz
3. Resample it to 32KHz (or apply a 16KHz lowpass filter?) and save as lossless file B
4. ABX file A against B
I'm not sure if replaygain and/or dithering should be apllied...
That's an interesting question.
There are some things to remember though:
1. Many lossy codecs are tuned for 44.1kHz, and either the tunings, or the fundamental basics of the encoder might not work as well at 32kHz (it implies more pre-echo in mp3 for example)
2. Sound cards aren't perfect. They might sound worse at 32kHz because they mess up, or because they force some horrid resampling, or because it brings all the faults around fs/2 down into the top of the audible range. You can solve this by using high quality resampling (e.g. using foobar2k), on playback to output at a sample rate at which your soundcard works well.
3. Ideal digital sampling implies a brick wall filter at fs/2 on conversion from and to analogue. If you can hear at 16kHz, the ringing might not sound nice, so it raises the question of what compromise to make.
Having said all that, I think it's worth trying. Apart from artificially generated test signals designed to stress D/A conversion, I've never been able to ABX 32kHz vs 44.1kHz with linear PCM - though by the time I learnt about ABXing, my ears were probably too old.
Did a quick test on giveuptheghost-sincealways and it was incredibly easy.
The lowpass is easily detectable.
ABC/HR Version 1.0, 6 May 2004
Testname: 44.1 -> 32khz resampling test
1L = giveuptheghost-sincealways.sample18sec.ssrc32.wav
1L File: giveuptheghost-sincealways.sample18sec.ssrc32.wav
1L Rating: 4.0
Original vs giveuptheghost-sincealways.sample18sec.ssrc32.wav
30 out of 38, pval < 0.001
The lowpass is easily detectable.
But you're young! (ish)
Yeah, my hearing dissapears above 16 kHz also. KikeG prepared some files here (http://www.kikeg.arrakis.es/lowpass/) just for this purpose: lowpassed real (well, somewhat manipulated, but real nonetheless) music v/s the original. I can ABX up to the 14 kHz lowpass, but not anything above that
The only test I remember was by ff123 on a favourite of his "peacful" by the eagles. It was a 15.5 lowpassed 32Khz resample.
He found improvement over the 44Khz version but still easy to abx.
dev0 what did you use for resampling?
From what 2bdecided says, I think that it would be better to keep files at 44KHz and apply some high quality lowpass filter. I also think that dithering shouldn't be used here. Still not sure about replaygaining...
Shame I had no time to do that test yet ...soon I'll try to prepare some samples for myself. I'm going to use SOX with "filter 0-16000" command for lowpassing.
fb2k's SSRC in 'slow mode' without dithering or rg.
Hmmm, using your way (fb2k's SSRC in Slow Mode w/o dithering & RG) I distinguished the sample 21 times on 21 tries (probability that I were guessing = 0%). That was easy indeed.
Using my way (files at 44KHz with lowpass performed by SOX using "filter 0-16000" command), I wasn't able to distinguish the files.
The sample I used was Garbage Collection by Wayfinder (music from FR14 intro by Farbrausch).
Could you please try to ABX again my way? And take a look at the spectrogram after using SOX, it might be that lowpass window is a bit too wide...
That's weird, cause SSRC's conversion is known to be 'as good as it gets' when it comes to sample rate conversion between standard samplerates.
Might be that one of the other factors mentioned by 2BDecided is causing the (perceived) quality degradation.
I'm currently using a M-Audio DMAN PCI Soundcard and Sennheiser HD-580 headphones (connected over some cheap HiFi amp). The DMAN PCI *does* resample, but it does a really good job at it (CS46xx chip).
I'll ABX a lowpassed sample tomorrow.
He was compressing to CBR 128 iirc. By improvement I think he meant the 32khz resample was harder to abx.