[music-dsp] OT: CD Perfection

[RTaylor]

"Angelo Farina" <farina at pcfarina.eng.unipr.it> wrote:

> Uhm...
> This intrigued me, so I performed a quick test with "noise-like" signal.
> I worked with Adobe Audition 1.5 and an external firewire soundcard: the
> Edirol FA-101, which is capable of 192 kHz too.
> So I started generating a piece of stereo white signal, at 192 kHz, 32 bits
> (float), long 9s.
> In the first part of the sound sample, I modulated the amplitude of the
> white noise, creating a transient effect, for three times - I have left the
> late part of the signal unprocessed (plain, continuous white noise).
> The spectral content of this white noise is flat (in a constant-percentage
> bandwidth analysis, such as FFT) up to the Nyquist frequency, which is 96
> kHz.
> After the first sound sample was prepared, I downsampled it with Audition
> (Edit - Convert Sample type), ensuring to employ the maximum quality
> (employing Pre/Post filter and moving the quality slider to 999).
> I created first a version at 96 kHz. Then I downsampled this to 48 kHz, to
> 44.1 kHz and even to 32 kHz.
> Looking at the waveform, it appears evident that halving the sampling rate
> destroys half of the energy (reduction by 3 dB), and the waveform appears of
> smaller amplitude on the screen. However, as we will see, it does not sound
> weaker...

More harsh. It may even sound louder.

> I am listening just now to the results, through a Sennheiser HD-560
> headphone connected directly to the headphones output of the Edirol. I am at
> home now, I do not have more professional equipment here, but I suppose,
> from what You said, that this equipment would suffice for hearing such an
> "evident" effect.
> The test is being performed single-blindly: my son is seating at the
> computer, playing at will the 5 sound samples (which for me appear as
> labelled A to E, and I do not know which is which, but I know if the one
> playing in each moment is A, or C, or whatever else).
> This test is NOT double blind (this would require that I do not know any
> reference to the sound samples, which makes the detection even more
> difficult).
> Perhaps later (after dinner) I would repeat the test double-blinded.
> The results here say that just a very, very subtle difference can be
> perceived when listening to the 32 kHz downsampled version; all the other
> are absolutely undistinguishable.

To me the difference is night and day. If not that... it's at least very evident... something that I don't think anyone would miss. I've tried this with every bit of software I own and every demo I've downloaded. This includes Audition, Soundforge, Wavelab, Audacity, Snd and Rezound and It's been sequenced in everything fro Cubase VST and Sonar{which I own... I've actually spent much time with them} to demos of Live, Samplitude and countless programs that I don't even remember the name of. The results are consistent across everything. 

Maybe I'm just more attuned to listening to noise but I really find it hard to believe that I'm more discerning or sensitive or whatever to that extent. It might be that I'm simply used to working with noise at higher hz and sensitive to the fact that I lose what seems to me to be major quality at lower rates. I really have trouble accepting that though. Maybe you should tune the samples to a slightly different frequency but I don't think that's necessary either.

Try playing the samples and watching. Maybe you just need to listen for the difference and the visual cue will help you.

> I have yet to see, however, if the difference for the 32k file is caused by
> the reduced frequency, or simply by the fact that I passed through the
> dowsampling subroutine for 4 consecutive times...
> I will retry, after dinner, with a 32 kHz version directly downsampled from
> 192 kHz....

Try not doing it in steps. Or just mess with it.

> In any case, I cannot second Your description of the problem from the test
> just performed. It seems that the downsampling algorithm of Audition is
> nearly perfect....
> But perhaps I did not employ the "right" signal for performing the test.
> In any case, I think that it is better if anyone else can repeat the test
> and we compare the results.
> So, I am posting the test waveforms on my web site.
> You can download them from here:
> HTTP://pcfarina.eng.unipr.it/Public/Downsampling
> As I am uploading from home, with an ADSL line, it will take about 20
> minutes to upload all of those 21.2 Mbytes...
> If You have a "better" test signal, originated at 192 (or, if not, at 96)
> kHz, Can You post it please?
> Bye!

I really don't. There really is no "right" signal. I could upload two files that sound different but I really think you just need to do it yourself to see the difference. It works just as well with plain white noise but I think maybe the degradation might be a little easier to hear if there's more variety.

I seriously just cannot spend any time at this right now. I am sorry. I don't mean to waste your time. I'm asking you not to waste it yourself. As I've said... I'm not really expecting a solution. If you're truly interested... go right ahead. :} I've spent years at it... I think it's just incredibly fascinating. That may not apply to you. I will try to download your files a bit later.