Quote:
Originally Posted by JaZZ /img/forum/go_quote.gif
Below a graph with a schematic illustration of high-frequency sine waves stored on a CD -- before low-pass filtering.
http://www.visaton.de:80/bilder/andere/sinuskurven.jpg
(...)
«No problem!» is the tenor of the Nyquist apologists. Because there's the classic «reconstruction filter» designed exactly for this purpose. Indeed: After being smoothed with the classic implementation of the anti-aliasing filter, the curves have turned into immaculate sine waves.
|
Exactly ! These pictures are misguiding advertisements in favor of high definition players. They illustrate some internal abstract mathematical way of dealing with data, which have nothing to do with the actual sinewaves that players output.
Quote:
Originally Posted by JaZZ /img/forum/go_quote.gif
But now imagine a scenario with an original signal catched by the microphone exactly corresponding to the amplitude-modulated sine wave above (and below). Again the «reconstruction filter» makes a continuous sine wave out of it. Correspondingly it makes the same -- smoothing and «(time-)smearing» -- with every other form of transients in an existing signal.
|
Exactly. The amplitude-modulated sine wave features a lot of frequencies above the cutoff frequency, that your graphs forget (on purpose ?), called aliasing.
A proper filter removes them, and let only pass the pure sine wave, which is exactly what the ear would do if one can hear up to 22 kHz.
The "reconstruction filter" is no more, no less than a lowpass.
In order to understand how it works, it is very interesting to play with sinewave generators, spectrum analysers, and lowpass filters in a software like SoundForge.
Adding two sine waves very close in frequency produces amplitude modulation. We can see only one sine wave in the temporal representation, but two close peaks in the frequency analysis.
Filtering out the higher one removes the modulation.
Quote:
Originally Posted by JaZZ /img/forum/go_quote.gif
Without a classic «reconstruction filter» (but a smoother slope instead) the response looks like this:
http://www.head-fi.org/forums/attach...1&d=1236602570
The measured high-frequency drop-off is the result of the amplitude modulation.
|
No. Amplitude modulation is caused by images of the audio band reflected above the cutoff frequency. Here is the complete frequency response (that commercials carefully stop at fc = 22050 Hz in order not to show the garbage that is above to their clients, and that is the cause of amplitude modulation) :
Quote:
Originally Posted by JaZZ /img/forum/go_quote.gif
Both phenomena -- HF drop-off and amplitude modulation -- show up in every Wadia player. Filterless DACs behave exactly the same, as they also renounce any form of FR reconstruction and AM smoothing.
|
Yes, and I have successfully ABXed (barely) the effect of this smooth loss of treble (ABX 7/8) while I can only hear up to 15 kHz !
On the other hand, I can't ABX high resolution vs low resolution (except for the bitrate : if I crank up the volume to 110 dB, in closed headphones, I have been able to ABX the quantisation noise of a 16 bits dithered version of a 24 bits original, during the initial fade-in of the sample).
Therefore, to my ears, checked under double blind conditions, the lack of anti-alias filter causes audible distortion, while lowering the sample rate from 96 to 44.1 kHz with anti-alias does not.
In order to get a realistic idea, the "Mustang" samples available on ff123's page are very interesting :
Samples for Testing Audio Codecs
They are lowpassed versions of the same sample, from 10 kHz to 19 kHz. They are extremely valuable, because they have been lowpassed using a Mathlab filter with a smooth, but short attenuation, like in ordinary DACs. We don't find this kind of filters in audio software (except maybe in resampling routines).
Therefore they keep nearly all their frequency content up the the cutoff frequency (usual audio lowpass starts the attenuation much below), while not ringing (since the filters have a smooth cosine profile instead of the brickwall, which causes very annoying ringing when it falls within the audible frequencies).
Personally, I can ABX the 13 kHz lowpassed one vs the original, but not the 14 kHz ! For this musical sample, a sample rate of 30 kHz would be enough for me (but I know that I could ABX higher cutoffs using pure sines, since I can hear them up and a bit above 15 kHz).
Quote:
Originally Posted by JaZZ /img/forum/go_quote.gif
To sum it up: «Reconstruction» in the term «reconstruction filter» merely addresses frequency response and completely ignores the time axis, thus impulse response. Nevertheless, some people still consider the transient corruption that comes with it inaudible -- for some reason --, a common reasoning is that the human hearing is relatively insensitive to transients, in contrast to frequency response issues. IMO this approach is quite arbitrary. A relative insensitivity -- even if it's true -- is still not the same as absolute insensitivity.
|
Yes, this argument is not convicing. The real reason is that if we look in the audio band, transient response is
completely unaffected ! All the time smearing that occurs is exclusively contained in the band where the filter operates (20 to 22 kHz) and above. That is why it is considered as unimportant. Because it is all above the audible frequencies.
Quote:
Originally Posted by JaZZ /img/forum/go_quote.gif
That said, the fact that there's a sonic difference at all also means that the ringing is perfectly audible -- despite the prevalent belief that it is not among technocratic circles.
|
You are comparing apples and oranges here. The assumption that there is no sonic differences only applies to blind listening tests. Everyone acknowledges that some differences are heard by some people in normal listening conditions.