you can't judge mp3 etc, they roll of much earlier due to their perceptual algorithm. Have a look into a spectrum analyzer! Also, a CD is rolling of REAL SHARP at 22k due to the Nyquist frequency. J On Apr 21, 2010, at 12:50 PM, Reuben Martin wrote:

On Tue, Apr 20, 2010 at 11:51 PM, Arnold Krille <arnold(a)arnoldarts.de>wrote;wrote: I agree, and this is something that is rarely talked about "scientifically" because the psychoacoustics of the phase information is so poorly understood. It's the spatial aspects of the sound that can be lost in the digital recording process, and that are carried in the phase information at the high frequencies. IMHO, compression psychoacoustics and human perceptual modelling also need to take into account how bats (or blind people) can echo-locate sound and come to an understanding of the physically-geometries surrounding them without traditional vision. Once that is better modeled, and the "internal deconvolution" that the brain accomplishes in deciphering what's echo versus source, what's "air" versus "ground' etc, one can then properly compress the "source sound" and then reconstruct the "air" to sound more realistic, without resorting to extremely high sampling rates (there's a reason 192K exists, and it's not just marketing). It is the muddling of our natural recognition processes that can make some recordings tiring to listen to. I think this "internal deconvolution" is also something that might be worth investigating as a better way of compressing sound. If there was a way of "reverse convolving" the echoed/heard sound back to the original source sound (given an impulse response of that instrument/track-->output); then one could use traditional MP3/OGG music encoding for the "original source sound" (where it won't matter if the phase&spatial information, esp at higher freqs, is lost). Concurrently with each frame of compressed audio, there'd also be a frame of "impulse response" that is separately coded and streamed. The "heard signal" is then recomposed from the decompressed MP3/OGG stream, and then convolved with a frame of "impulse response". (The smooth inbetweening of convolutions between frames of impulse response is left as an exercise for the reader :-) ). -- Niels http://nielsmayer.com PS: I think learning how MP3 works is like finding out what went into the sausage you previously thought so delicious... might even make you go vegan. What's interesting is that all the stuff on psychoacoustic modelling is presented as a "total fact" with no consideration for any of the "perspective" raised above. For example, the only place "time domain" is mentioned is in the unverified assumptions about "temporal masking":: http://www1.cs.columbia.edu/6181/slides/05-music-coding.pdf http://www.cs.rutgers.edu/~elgammal/classes/cs443/slide14_short.pdf ... IMHO, the literature on compression ends up reading like medical literature where they've already decided what medicine they're going to sell you, then devise tests, diagnoses, and even new diseases to justify the "sale", while purposefully ignoring (as part of their misuse of the scientific method) anything that doesn't fit with the idea they're desperately trying to sell...

On Wed, Apr 21, 2010 at 01:47:32PM -0500, micromoog wrote: I found the links privided by Niels Mayer quite useful on this topic: http://en.wikipedia.org/wiki/Nyquist_frequency#The_aliasing_problem Based on my own experience, it certainly happens, though I can't say why, it may be resampling or crappy hardware as others pointed, but the distortion is there and audible. It becomes impossible to notice with high sampling rate recordings (at least on PC hardware, have not tried on portable players yet). Maybe it can be avoided somehow, but disk space is cheap nowadays so I find more convenient to just record at 96Khz and avoid headaches.

On Wed, Apr 21, 2010 at 4:34 PM, Rob <lau(a)kudla.org> wrote: In any good system, both are going to be silence, as they've been entirely lowpassed away. Choosing something a little lower (like 18kHz), they'll still be identical, as the harmonics that make a square wave a square wave will also have been lowpassed away. The square wave's first harmonic (OK, it's actually called the 'second harmonic, as the 'first harmonic' is the fundamental) is at 36kHz, above Nyquist. When you filter away all the harmonics of a square wave, what do you get? A sine wave. When you filter away all the harmonics of a triangle wave, what do you get? A sine wave. Completely incorrect. No shred of truth, period. I grant this is an advanced a topic that relatively few people have studied. But just because you personally don't understand it doesn't make it 'not so'. Monty

Speaking for Vorbis, it depends on the quality mode or bitrate you select. Vorbis is also fully aware ot the ATH and absolute dynamic depth of the human ear. If its psychoacoustics say a tone is inaudble, it won't code it. Energy is preserved, even at the exact nyquist frequency, however modulating the input signal against the sampling clock results in the signal being partly shifted to DC. The signal energy is there... it appears somewhere else. This is the case only with a signal exactly at Nyquist, unless you also have a system exhibiting IMD (most modern ADCs and DACs are very very clean. I can't measure the IMD even on my ten year old samplers, though they're much noisier than today's). The aliasing caused by signals exceeding the nyquist and folding back over into the sampled spectrum as well as aliasing caused by IMD at or near the sampling frequency are both mitigated/practically eliminated by lowpassing somewhere south of Nyquist. This is utter hogwash. All the signal energy and resolution is preserved to exactly the limits of the sampling rate at all lower frequencies, even if your 'biggest dots' aren't landing at 0dB or wherever. True of audio signals, true of images, true of video. The same discrete sampling lessons apply to all three equally. Monty

On Wed, Apr 21, 2010 at 10:08:22PM +0200, fons(a)kokkinizita.net wrote: Heh :) It happens on all topics, not only acoustics. The real fights start when someone tries to scientifically demonstrate creationism :P Anyways, going back to the original topic, no matter if DAC, resampling, bad software, analog amplifiers or whatever the problem is, a lot of consumer devices distort when playing high frequencies (I think everybody agree on this). It may be the reason for some recordings start filtering a bit before 20Khz... It is weird that my Delta 66 also seems to distort at 44100 (though much less noticeable) and is perfect at 96000, but it may be completely different reasons I would say.

Monty Montgomery <xiphmont(a)gmail.com>om>: modulating the input signal against the sampling clock results in the signal being partly shifted to DC. The signal energy is there... it appears somewhere else. I bet a scientist could come up with a wacky experiment to disprove your assertion that nyquist is the only thing that mattered. For example, assuming one had a way of "playing back" in real-time, echoes of what a bat would hear as it flew around a room of known geometry.... Consider an experiment: Record/playback one using the "pure nyquist" hypothesis, and therefore containing only "energy" or real-plane information. Record/playback another assuming "nyquist++" -- nyquist theory augmented by hypothesis that both real and imaginary plane information can be perceived by mammals with advanced hearing capabilities (dogs, bats, humans, dolphins). Do the bats virtually "fly into walls" when presented with "pure nyquist" virtual audio world? Or do they navigate correctly when presented with sampled/reconstructed representations that had concern over both real and imaginary plane info? (aka phase matters, esp at high frequencies, and therefore you need much higher sampling rates). Note that the experimental subject, being a bat, is probably unbiased to the argument at hand. Further, no double-blinding is needed because, again, they're bats :-). This is utter hogwash. All the signal energy and resolution is preserved to exactly the limits of the sampling rate at all lower frequencies, even if your 'biggest dots' aren't landing at 0dB or wherever. True of audio signals, true of images, true of video. The same discrete sampling lessons apply to all three equally. Prove it! I don't need another paper telling me about nyquist or how compression works. Until you have a proper model of human perception, this kind of cocky blinded-pseudo-rationalism, or repeating the same old tired truisms -- is just a waste of everybody's time. Nyquist is at best a sophomoric model of the *human* perception of sound, the biological perception, of the "energies" you're talking about. It's like Marie Curie saying "well it only fogs the film"... yeah, and in biological organisms it causes cancer -- because it's not just the energy, it's also the frequency (and in this argument, phase). Furthermore, your perspective totally ignores the fractal nature of reality -- of nature. And of sound. Fractal image compression is resolution independent and works precisely because it better models real-plane and intra-dimensional aspects of reality that you've decided to rationalize away. You claim this is true for images and video too: How does it work in fractal compression? ( http://en.wikipedia.org/wiki/Fractal_compression ). And when you think about it fractals, and not "masking" is the correct way to compress sound, as it's most likely directly involved in our ability to perceive sound: Ultimately, the brain isn't masking anything, it's taking each instrument heard and "triggering" the fractal-equation that would generate it's sound. So if we perceive sound as a bunch of activated fractal recognizers -- the fractal that generates the "bass" sound... the fractal that generates the "drum" --- our brain reconstructs the missing pieces no matter what. Just like a broken hologram can still reconstruct the original image. See also Michael Barnsley's "Fractals Everywhere" (it's on my nightstand now, alongside David Wright's "Mathematics and Music"). Or one of the most mind-blowing papers ever: http://www.ams.org/notices/201001/rtx100100010p.pdf "The Life and Survival of Mathematical Ideas" Michael F. Barnsley. (and also in the same issue http://www.ams.org/notices/201001/rtx100100030p.pdf ) Niels http://nielsmayer.com

Then I have no hope. and prefer not to try. Monty

On Wednesday 21 April 2010 08:51:22 Arnold Krille wrote: Okay, the answers teach me that this is wrong. I didn't yet think about this as in-depth as I should have before writing that. Thanks for the clarifications, Arnold

On Thu, Apr 22, 2010 at 12:19:27AM +0200, Arnold Krille wrote: Well done !! This is how you learn and understand things. The 'amplitude modulation' is an artefact of looking at a graphical interpretation that doesn't do antialising. It's the sum of f=0.49 and f=0.51. When the latter is removed, the 'amplitude modulation' disappears. Right. For f=FS/2 the alias and the original are the same except that the alias has the imaginary 'sine' part inverted (by symmetry). So the 'cosine' part (the one where the peaks of the waveform coincide with the samples) is doubled, while the 'sine' part (samples at zero crossings) gets cancelled. Ciao, -- FA O tu, che porte, correndo si ? E guerra e morte !

On Wed, Apr 21, 2010 at 10:28:06AM +0200, fons(a)kokkinizita.net wrote: Hello Fons. Almost all consumer portable devices seem to distort signals above 16Khz in uncompressed 44100Hz audio. I always wondered about this, maybe they all have very crappy DA converters, or maybe it is because the reasons explained by Niels and Arnold in their emails... no idea, but the distortion is real and present in the vast mayority of devices people commonly use to play music nowadays, sadly enough. The most common artifact is the addition of frequencies not present in the original. I can barely hear them (and I guess most people won't care), but I know a person who was extremely annoyed by this fact; he is a musician and has very good trained ear, so I ran several tests on a wide range of media players, PDA, high end mobile phones, portable tablets and such devices trying to find something that is not that crappy. Funny enough, the distorsion is usually not present in compressed mp3 audio, maybe because the compression itself filters high frequency signals so they are not an issue anymore. Nowadays, people (specially young people) even listen to music using the tiny external speaker built in their mobile phone (which can barelly reproduce spoken voice) and they don't care about the quality. It seems that companies feel no reason to improve their crappy sound hardware :( But yes, agree, please don't destroy your music just because of this. Most people who don't care will convert to 128kbps mp3 anyways, so please keep your music as good as possible for the rest of us ;)

On Wed, Apr 21, 2010 at 05:24:37PM +0200, Arnold Krille wrote: Interesting. I should test that. Do you think that it can also happen when downsampling (for example playing 96Khz wav files)?

It is not the DACs, it is almost always the analog amplifiers. Have a look at the various noise and distortion measures of audio opamps, the figures tend to dogleg upward between 10 and 20kHz. It's a consequence of optimizing for very very deep signal depth below those frequencies (and to a lesser extent, cost). Even the most expensive boutique audio opamps show this dogleg, and very few manage to push it higher than 20-25kHz or so. To do so, you have to sacrifice noise performance. You can get opamps that are dead flat with unmeasurable distortion to 100kHz or 1MHz... but you're only going to have 8 bits or less of signal depth due to voltage and current noise. Monty