Linear PCM is crap, SACD is best

Topic: Linear PCM is crap, SACD is best (Read 37363 times) previous topic - next topic

0 Members and 1 Guest are viewing this topic.

Linear PCM is crap, SACD is best

Reply #25 – 2004-04-01 17:59:26

Quote

This is so basic that I would be extremely surprised if any of the posters in this thread didn't already know it.

Maybe they know it but don't have a feeling for it. Most of the time signals are quite low in music, especially for the parts where the pinpointing is important. They are incredibly low and it is important to sample these siganls with high enough resolution. In 16 bits systems this is clearly not done and in 24 bits it is considerably better. Also in the range of 3500 Hz, where our hearing is best, are these signals small.

Quote

That page doesn't show any more than that if you quantize to 24 bits a signal with a peak that's below 18 bits, the signal has an effective resolution of 6 bits.

which is exactly the technical reason to develop a different digital standard in such a way that this signal has also >20 bits of resolution. Not forget that DSD was developed especially for systems where these small variations must be stored with the highest possible quality. A commercial application appears to be SACD.

If one analyses classical music, one will be surprised how much low-energy signals are present in audio.

cheers,
Jacco

Linear PCM is crap, SACD is best

Reply #26 – 2004-04-01 18:07:48

Quote

They are incredibly low and it is important to sample these siganls with high enough resolution.

Why? You cannot possibly hear any noise resuling from the quantization anyway.

You still didn't answer my question: how can a dynamic range that is large enough that when the top makes you instantly deaf, still has a noise floor below the audibility threshold, be possibly 'not enough'?

Linear PCM is crap, SACD is best

Reply #27 – 2004-04-01 18:55:33

Quote

Why? You cannot possibly hear any noise resuling from the quantization anyway.

I don't understand your comment. The quantization noise cannot be heard with 24 or 16 bits I agree. But it is vital to have high resolution in the high frequency range. Let me show an example I just did. I took a classical piece with violins and stuph and it was not a loud passage because further on the CD there was a fortissimo and that sets the peak value. The average of this is at approx -20 dBFS, which is not bad for a good recording. It could be ten dB higher if there was not a very loud passage furtheron, but this is the choice of the recording engineer. The result can be found here. (right now I have some problems with the server)

The measurement is done with peakhold enabled, so the levels you will see are peak levels and it can be seen that the low frequent peak is at about -40 dBFS and in the range of 10 kHz, the signal is around -80 to -100 dBFS. Comes close to the -108 dBFS level which Daisy-laser has used. This shows clearly that 16 bits is not enough and 24 bits is better but probably not enough.

Quote

how can a dynamic range that is large enough that when the top makes you instantly deaf, still has a noise floor below the audibility threshold, be possibly 'not enough'?

Because we human beings turn the volume knob accordingly. In loud passages we turn it back (especially with boom-boom pieces) and if we want to listen to classical music, the average loud-level is around -10 dBFS and (appearently) the peaks of low-level are at -40 dBFS and the real small ones at -90 dBFS. This is with an overall level of -20 dBFS.

By the way, the frequencies that make you deaf are at relatively low frequencies. A tweeter of a loadspeaker gets burned alive if you provide it with more than 10 (electrical) Watts.

Linear PCM is crap, SACD is best

Reply #28 – 2004-04-01 19:22:36

Quote

The measurement is done with peakhold enabled, so the levels you will see are peak levels and it can be seen that the low frequent peak is at about -40 dBFS and in the range of 10 kHz, the signal is around -80 to -100 dBFS.

I don't see the graph, but what exactly says that that 10kHz -100dB signal is even audible? Let alone the quantization noise *another* 40dB below that?

In the link you posted earlier, they try to pull a similar trick by zooming in and showing how 'rectangular' the signal now looks. But that doesn't mean the difference between the two is audible.

Quote

Because we human beings turn the volume knob accordingly.

I certainly don't agree with that assumption, i.e. that a listener will be turning the volume knob to an extent that -140dB quantization noise is audible *during* a music piece.

To give an example, imagine that the level is indeed at -90dB and you turn that up to get a normal playback level. The next movement begins and is at full scale again. Painfull experience.

In the example you give you yourself notice that the mastering engineer stopped at 20dB volume differences for this very reason.

I do believe we compensate for differences between tracks (cue ReplayGain), but these are not of the order you are discussing here. Top differences are also around 20dB. You'd _still_ not be close to hearing the quantization noise with the very quietest DVDA.

Quote

By the way, the frequencies that make you deaf are at relatively low frequencies. A tweeter of a loadspeaker gets burned alive if you provide it with more than 10 (electrical) Watts.

You can replace 'makes you deaf instantly' by 'makes your speakers go up in flames' in my posts then. Not that it helps your argument.

Linear PCM is crap, SACD is best

Reply #29 – 2004-04-01 19:37:19

Quote

don't see the graph, but what exactly says that that 10kHz -100dB signal is even audible? Let alone the quantization noise *another* 40dB below that?

I sent it to you by e-mail. And it is audible, because I hear (when I listen to the music) frequencies above let's say 1 kHz. So these small levels must be audible...

Quote

But that doesn't mean the difference between the two is audible.

You are right, there is no listening test mentioned there but the resulting "wave" is very different from the original and therefore was this digital format rejected as storage-technology and is choosen for a different format, ie. DSD.

Quote

a listener will be turning the volume knob to an extent that -140dB quantization noise is audible *during* a music piece

You will never hear -140 dBFS since it is much lower than the noise in the system. A common amplifier has only 75 dB SNR and is just good enough to reproduce the classical piece which I just examined. But if you listen to this, you will have set the volume control to a higher level. otherwise it is pretty silent in the room. Furthermore, you cannot improve the sound quality by increasing the volume since the damage is already done. So you will listen to distorted high frequency content, for what it is worth.

cheers and I am going home,
Jacco

Linear PCM is crap, SACD is best

Reply #30 – 2004-04-01 19:44:03

jdekkers picture

Linear PCM is crap, SACD is best

Reply #31 – 2004-04-01 19:57:18

I am really starting to wonder if this guy is for real. He seems suprisingly ignorant of basic topics in signal processing for somebody who worked in a digital audio research lab.

Quote

You are right, there is no listening test mentioned there but the resulting "wave" is very different from the original

No. It isn't. Haven't you been listening to what Garf has to say? The resulting wave of a LPCM sampling process has a limited bandwidth (Fs/2) and raised noise floor (due to sample quantisation). That is all. However different the waves may look, these are the only differences. Whether or not they are audible depends on whether you can hear the band limiting (not possible with 96kHz sampling) or whether you can hear the noise. As Garf has said, the noise floor in 24 bit PCM caused by the digitising process is incredibly low. That is, DVD-A's amplitude resolution is much greater than that of our ears.

I could provide strong scientific proof for all my and Garf's statements. Could you? Dekkersj, either you need to provide proof for your claims or you need to admit defeat. I know my Signals professor would be pretty suprised by your proof, as would most people on this board.

Linear PCM is crap, SACD is best

Reply #32 – 2004-04-01 20:03:49

Quote

I sent it to you by e-mail. And it is audible, because I hear (when I listen to the music) frequencies above let's say 1 kHz. So these small levels must be audible...

*cough* The level in the spectograph doesn't go below -100dB until way over 10kHz, so quite obviously there are audible things over 1kHz...

Quote

You are right, there is no listening test mentioned there but the resulting "wave" is very different from the original and therefore was this digital format rejected as storage-technology and is choosen for a different format, ie. DSD.

The way the wave LOOKS is an argument now? Even if we take the discussion to something as peverted as that, how on earth is a 2.8Mhz 1 bit sampled signal going to look CLOSER to the original?

Even on playback this holds no water as DVDA still has the better characteristics. I'm quite sure that what comes out of your SACD won't bear much resemblance to what went in, especially at those high frequencies you were just ealier claiming to be so important. The entire working of DSD is based on it.

Quote

You will never hear -140 dBFS since it is much lower than the noise in the system. A common amplifier has only 75 dB SNR and is just good enough to reproduce the classical piece which I just examined. But if you listen to this, you will have set the volume control to a higher level. otherwise it is pretty silent in the room.

Yes. And your volume control is independant of the level of the CD. If the CD peaks at -40dB only, the mastering was fucked up. Now, you can go and fuck up DVDA's and SACD's too, but that will not tell anything about the capabilities of the format.

That being said, again using your example, the peak being at -40dB STILL gives a 100dB dynamic range left. Bigger than what your equipment can handle, as you so nicely pointed out!

Linear PCM is crap, SACD is best

Reply #33 – 2004-04-01 21:07:38

Quote

Now, you can go and fuck up DVDA's and SACD's too, but that will not tell anything about the capabilities of the format.

That is the real problem, the inevitable radio loudness creep over time, not that 24 bits does not have the resolution, (well that and the fact a DVDA disc costs 2x a CD does, or as much as 4x as a video DVD!).

Linear PCM is crap, SACD is best

Reply #34 – 2004-04-01 22:18:49

We have been debating the relative merits / lack thereof of DSD/SACD. Going back to the OT of this thread - how about "LPCM is crap?"

dekkersj, can you please explain why LPCM is "crap?" It would be hard to understand how a 24/192 DVD-A signal could be "crap." We are talking about a 144dB dynamic range from DC ~ 96kHz for goodness sake, and with some noise shaping it could easily achieve 160dB+ of dynamic range in the audible frequencies. There has never been hardware that could even approach this dynamic range for recording or playing back audio. I submit that 24/192 LPCM is an extremely transparent format for audio.

How could 160dB+ dynamic range and a frequency resopnse flat from DC ~ 96kHz be "crap?"

Linear PCM is crap, SACD is best

Reply #35 – 2004-04-01 23:51:04

Jacco, it would be much more useful if you could do a blind test. If LPCM is rubbish, even at 24/192, then surely 16/44.1 must feature the same problems, but far more prominently, right? So you should be able to easily tell them apart, if what you are saying is true..

Linear PCM is crap, SACD is best

Reply #36 – 2004-04-01 23:55:36

Levels in spectrum analysis can't be compared to levels of the wave form.
For example I just analyzed a wav peaking at -1 db, with an RMS level of -16 db. 44.1 kHz 16 bits, completely random.
Not a single frequency is above -36 db in the spectrum analysis ! But it is obvious, given the peak level, that all these frequencies were recorded with at least 15 bits accuracy (peak between -6 and 0 db), while following your argument, they would only have 10 bits of accuracy.

Linear PCM is crap, SACD is best

Reply #37 – 2004-04-02 08:18:08

Never mind that this guy has never provided an argument for why DSD would do better capturing these low level signals.

Linear PCM is crap, SACD is best

Reply #38 – 2004-04-02 09:10:49

When you think the claims presented here are troll tales you have to read this silly thread at diyaudio.com
http://www.diyaudio.com/forums/showthread....15&pagenumber=1
I wonder why the admins over there aren´t a bit more consequent and "Do the Garf" a bit

Wombat

Linear PCM is crap, SACD is best

Reply #39 – 2004-04-02 09:35:10

Pio2001, you obviously found the flaw in dekkersj's reasoning. Some additions:

If you reapat your test (fft analysis of random noise) with different window lenghts you'll notice that the total energy (which is constant) is split between fft frequencies, so for increasing window lenght (number of analysed samples) the amplitude/energy becomes lower. So if you choose a big enough window lenght, the amplitude for each frequency will be lower than the quantization noise for your bitrate of choice. This should be a 1st clear hint that drawing conclusions from comparing both values is flawed.

Example: I created a 1000->5000Hz sine sweep at 44.1kHz/32bit float with CoolEditPro, changed resolution to 16 and 24 bit using flat and noiseshaped dither and performed frequency analysis 2 times; one time with 128 samples fft size, the other time with 65536 samples:

Linear PCM is crap, SACD is best

Reply #40 – 2004-04-02 11:33:34

Quote

Any need for explanation left?

Cool Edit Pro spectral analysis isn't very good?

Seriously - with a given FFT length and window, it's helpful to calibrate the dB display so that 0dB equals the energy from a digital full scale sine wave centered on one FFT band. (This leaves the theoretical possibility of signals over 0dB in the spectral view, but it makes more sense than no calibration at all!)

I don't know what happens in CEP - it seems it's calibrated like this for some combination of FFT length and window function, but not for any others.

In other words, most of the time, those dB values don't mean anything. dB is always relative, but exactly what it's relative to isn't clear with CEP spectral analysis.

Cheers,
David.
(P.S. another reply on the way - that link is so stupid!)

Linear PCM is crap, SACD is best

Reply #41 – 2004-04-02 11:50:15

Quote

Why is DSD needed?

OK, first of all...

It's easy to have two time-domain waveform graphs that look identical, but sound different. You just hide some spectral or temporal vs amplitude feature just below the 1 pixel resolution of the display.

It's easy to have two time-domain waveform graphs that look completely different, but sound identical. If you add a full scale sine wave at 1GHz, it'll change the display pretty dramatically, but I don't think you'll hear it!

Secondly...

(and Garf beat me to it, but let's say it again...) what do you think an SACD waveform will look like when you sample -108dB noise? Answer: almost the same as every other SACD waveform - a 2MHz digital full scale square wave!

If you filter out the ultrasonic hash, it can't look any closer to the original than a 24-bit 192kHz version, so this pro-SACD argument is useless from all angles!

Thirdly...

For all the reasons why, technically, SACD sucks, please see...

Why Professional 1-Bit Sigma-Delta Conversion is a Bad Idea
Why 1-Bit Sigma-Delta Conversion is Unsuitable for High-Quality Applications
Towards a Better Understanding of 1-Bit Sigma-Delta Modulators
Towards a Better Understanding of 1-Bit Sigma-Delta Modulators - Part 2
Towards a Better Understanding of 1-Bit Sigma-Delta Modulators - Part 3

All by Stanley Lipshitz and John Vanderkooy. All presented at AES conventions between 2000 and 2002.

(I'm aware of the Philips sponsored responses from James Angus - he was my external PhD examiner - but I don't buy his arguments, and neither do Lipshitz and Vanderkooy, so I'm in good company ).

Beyond this, reasons why SACD and/or DVD-A may sound better than CD (and, more basically, proof that they actually do) are always welcome.

I personally believe that 96kHz sampling sounds better than 44.1kHz sampling, and that SACD does sound different from either. I have neither proof, nor explanations. I've posted my best guesses plenty of times already - now it's your turn!

Cheers,
David.

Linear PCM is crap, SACD is best

Reply #42 – 2004-04-02 11:59:43

Quote

I am really starting to wonder if this guy is for real.

Yes, I am real but also a guy who doesn't care to row against streams.

Quote

suprisingly ignorant of basic topics

Well, a basic topic in LPCM is bitresolution and that is (and only that) my concern. This is an example of rowing against streams. I wil not debate if the Dynamic Range is large enough, because that is the case already in 16bit/44k1. Dynamic Range is more than enough, apractical number of a DA used in audio is 117 dB for a 24 bits system and this much more than enough.

Quote

The resulting wave of a LPCM sampling process has a limited bandwidth (Fs/2) and raised noise floor (due to sample quantisation). That is all.

If one applies the sampling theorem and assumes that an infinite number of bits is used (NOT TO CREATE Dynamic Range, but to be able to reconstruct the original waveform) and ideal filtering is applied. Well, the latter constraint is obviously reached in 192 or 96kHz systems, or in oversampled systems. To emphasis it one more time, my (and a lot of other researchers oround here) problem with LPCM is the limited number of bits to sample low-level signals. Large signals are out of the question, here is LPCM good enough.

Quote

Signals professor

He also knows that low-level content is heavily distorted in LPCM. Here, where the CD is invented, the assumption was made that the probabilty of those low-level content is that low that it could be ignored. At that time, the number of bits was limited and these two arguments form the basis for the CD as it is. Very soon after the introduction of the CD this

Quote

flaw

was "recognized" by high-end audio people. I know of one small company (because I worked there) which was sewed in court by trying to spread this word. As a consequence this small company had to be silent for 3 years. They also row against streams.

Quote

Even on playback this holds no water as DVDA still has the better characteristics. I'm quite sure that what comes out of your SACD won't bear much resemblance to what went in, especially at those high frequencies you were just ealier claiming to be so important. The entire working of DSD is based on it.

Next week I will try to get some examples of such a wave, sampled with DSD technology. Basicaly, the working principle of DSD is as follows: you have a comparator and this will compare the input signal with a sawtooth. As long as the signal is above the hysterese of this comparator, it will detect the input signal. And since the frequency is insane high (2.8MHz) every subtle change at any level will be detected. This also causes a lot of ultrasonic crap and this is a drawback. Modern systems however can accomodate with that and it is doable. Research can help to think of systems to reduce these artifacts.

Quote

We are talking about a 144dB dynamic range from DC ~ 96kHz for goodness sake

Yes, but this at his best (the 144 dB) and it shows clearly that many peolpe see in Dynamic Range the answer to all their problems. Today, in consumer DAC's the DSD and 24bits LPCM modes has a Dynamic Range of 114 and 117 dB respectively (CS4398). This can be a political statement of coarse, but I do believe it since there is also some electronics involved after the DA itself.

Another "proof" can be found in a completely different type of application. In our lab we deal with systems where modulation and demodulation is done with I/Q signals. Has nothing to do with audio, but here is where the limitation of LPCM comes in. I and Q signals (bandpass representations of an RF signal) must be exactly 90 degrees out of fase, otherwise demodulation is crapped. To test a transceiver you need an I/Q signal generation and these signals must have excellent quality, because degration of your DeviceUnderTest is to be investigated. One additional BNC connector in the leads towards the I/Q modulator is enough to spoil a measurement, to give an idea of how exact things must be. It is found that if the digital stream is not with full resolution, no measurement can be done. The setup is: digital generated file->DAC->I/Q modulator. Experiment proves that if one uses signals below 0 dBFS, the mismatch between the signals is (slightly) different than 90 degrees. The lower the signal the more mismatch. So this must be rejected:

Quote

That is all. However different the waves may look, these are the only differences.

since there is also a difference in time (phase) and/or amplitude...

Oversampling is used to minimize this problem, not to overcome.

cheers,
Jacco

Linear PCM is crap, SACD is best

Reply #43 – 2004-04-02 12:02:51

Quote

I am really starting to wonder if this guy is for real. He seems suprisingly ignorant of basic topics in signal processing for somebody who worked in a digital audio research lab.

Well, look at the date could it be April Fools?

Else he's mixing up resolution, bitdepth and signal strength (loudness) on more than one occasion.

(and of course as Joe said: no proof for the supposed superiority of DSD).

Linear PCM is crap, SACD is best

Reply #44 – 2004-04-02 12:47:43

Quote

Basicaly, the working principle of DSD is as follows: you have a comparator and this will compare the input signal with a sawtooth.

That's simply not true! You're confusing DSD and Class-D again.

Quote

Another "proof" can be found in a completely different type of application. In our lab we deal with systems where modulation and demodulation is done with I/Q signals. Has nothing to do with audio, but here is where the limitation of LPCM comes in. I and Q signals (bandpass representations of an RF signal) must be exactly 90 degrees out of fase, otherwise demodulation is crapped. To test a transceiver you need an I/Q signal generation and these signals must have excellent quality, because degration of your DeviceUnderTest is to be investigated. One additional BNC connector in the leads towards the I/Q modulator is enough to spoil a measurement, to give an idea of how exact things must be. It is found that if the digital stream is not with full resolution, no measurement can be done. The setup is: digital generated file->DAC->I/Q modulator. Experiment proves that if one uses signals below 0 dBFS, the mismatch between the signals is (slightly) different than 90 degrees. The lower the signal the more mismatch. So this must be rejected:
Quote
That is all. However different the waves may look, these are the only differences.
since there is also a difference in time (phase) and/or amplitude...

Oversampling is used to minimize this problem, not to overcome.

With correct dither and good filters, the relative phase of 2 different 1kHz sine waves (e.g. one on each stereo channel) is perfectly preserved (subject to known uncertainty principles) in a 3kHz (!!!!) sampled system. It is not lost at less than full scale with correct dither (subjective to known uncertainty principles - if there's more noise than signal then you have a problem - if there's more signal than noise, even 80dB below full scale, then the phase is preserved).

If the D>A converters destroy this, then that's the fault of the converters. The information is present in a correctly dithered Nyquist sampled digital audio signal; good converters can recover this information.

When it comes to phase and timing information, bad converters don't mean you need a higher data sample rate, they mean you need better converters!

Cheers,
David.

Linear PCM is crap, SACD is best

Reply #45 – 2004-04-02 12:50:40

@ dekkersj:

who in their right mind is going to hook RF into their amplifier and try to listen to it?

we're talking about AUDIO here. i seriously doubt whether you (or a golden eared research subject) could even ABX DTS and LPCM, let alone identify a phase difference of perhaps 0.001 degrees.

stay on topic. you first say LPCM doesn't represent transients right, then after that was disproved you say the problem is frequency, THEN you're saying the problem is phase.

if this is an example of sony-philips research, then i'm not going to be replacing my CD player in a hurry... sorry. (yes, i know what company developed the CD player. maybe their best work is behind them?).

Linear PCM is crap, SACD is best

Reply #46 – 2004-04-02 13:50:25

Quote

Well, a basic topic in LPCM is bitresolution and that is (and only that) my concern. This is an example of rowing against streams. I wil not debate if the Dynamic Range is large enough, because that is the case already in 16bit/44k1. Dynamic Range is more than enough, apractical number of a DA used in audio is 117 dB for a 24 bits system and this much more than enough.

Let me check if I understand correctly what you're saying:

- Dynamic range increases with bit depth (16 -> 24bit), "bitresolution" does too, but it's still not enough to represent (some) signals audibly transparent (audibly transparent means there's no difference audible between (analog or with higher bit depth sampled) original and sampled digital version, both played back with state-of-the-art equipment).

- Even signals with peak values like -20dBFS can be regarded as a combination of many low level sine waves (with e.g. -80dB peak level). The bits left at such low levels aren't enough to represent the sampling points of these sine waves exactly enough, and this is the too low "bitresolution" you're talking about.

Please confirm if I understood this correctly or correct me.

In case I got you right, can you please tell in what way this too low bitresolution affects the perceived audible quality in your opinion? I.e. do you expect audible noise added, distortion (= harmonics of the original frequency added resulting in the sound being "coloured"), discrete steps on volume changes (= e.g. fadeouts being not continuous but consisting of constant volume steps), or something else?

Linear PCM is crap, SACD is best

Reply #47 – 2004-04-04 11:57:21

Tigre,

Quote

- Dynamic range increases with ... state-of-the-art equipment).

Yes, I confirm that this is the case. A rock solid evidence (for me) is the example I gave of the two signals which must be out of fase. Right now I am studying this in order to reproduce it with mathematical software which can be repeated by anyone. I admit, those differences are small but I don't know how small. Also the influence of oversampling and dither is interesting, because it is clear that this is necessary since the basic technology is incapable of transparent signals. Interesting question: how can dither improve a DC signal? Answer: Not, you can dither a DC signal whatever you want, it will not improve. Next interesting question: how can dither improve very low to low frequencies (in signals) improve the digital representation? Overall question: what is the optimum dither to get a good comprimise for all frequencies?

There was also a remark about levels that could not be compared in the "wave"domain and frequency domain. As far as I know they are linearly interrelated (time and frequency domain) by means of Fourier theory. Any wave can be represented as a superposition of signals like (1+A)cos(omega*t+phi), where it is clear that all properties are of importance.

Quote

stay on topic. you first say LPCM doesn't represent transients right, then after that was disproved you say the problem is frequency, THEN you're saying the problem is phase.

So, if a transient isn't represented right (and I am only considering the small ones, since large transients will be sampled good enough), there are problems with amplitude (obvious), and/or phase. The higher the frequency to be represented the larger this problem will be, assuming lower signal amplitude at higher frequencies.

Quote

Even signals with peak values like -20dBFS can be regarded as a combination of many low level sine waves (with e.g. -80dB peak level). The bits left at such low levels aren't enough to represent the sampling points of these sine waves exactly enough, and this is the too low "bitresolution" you're talking about.

Yes. Those musical signals have also harmonics in the time domain and they will be much smaller at higher frequencies. One can represent that as superpositions in the time domain as well (addition is a linear operation and can be seperated from the Fourier operation). The relatively large zero harmonic is well enough represented, the first harmonic somewhat less and so on. Or if you want to "see" it in the time domain completely, you will have a signal with large amplitude and a lot of strange looking deviations from that (these are the harmonics). Deviations smaller than one bitresolution-step will not be detected, but this is obviously the noisefloor. The larger these deviations become (ie. the lower the number of the harmonic) the better they will be described relatively to their amplitude in the frequency domain by the digital word that comes out of the AD. My concern is not how to make it perfect, since I believe that that is impossible. Whatever the number of bits you have, you will have per bit a defined level at which the bit turns from 0 to 1. Given this, one can say that the probabilty of exact this value is 0 (zero) in any continues stochastic process which has a power-spectrum-density. At least in mathematics this is true.

Quote

low bitresolution affects the perceived audible quality in your opinion?

The first thing what pops up is stereo image. Distorsion is very likely (can be in the frequency and/or phase domain as well as in the amplitude), but I will try to investigate this more with mathematical tools.

cheers,
Jacco

Linear PCM is crap, SACD is best

Reply #48 – 2004-04-04 14:51:57

Quote

how can dither improve a DC signal? Answer: Not, you can dither a DC signal whatever you want, it will not improve.

ummm....

DC sounds just as quiet in PCM no matter how many bits you throw at it and no matter what sampling rate. silence is silence. i don't see what baring this has on your argument (maybe i'm thick - after all i don't work in R&D, but am merely an interested observer)

if you were to dither silence, what would the outcome be? adding a random (or noise-shaped) element to zero would still give zero, plus maybe some hiss. same with oversampling... average of 4 zeros is zero.

interesting to me is that at some point DSD and LPCM can be considered equal. try this in Audition or similar: get a sample sound (i tried with a chunk out of "mars - the bringer of war", but it'll work with anything). do a high quality resample to a stupidly huge rate (my computer nearly choked, but i went up to 882,000 hz). then change the bit depth to, say 4. make sure you use noise shaping.

now... convert back to a high bit-depth, lowpass at the original cutoff, and resample back to the original rate (mine was 44100).

you'll notice that there isn't much change in the signal.

what you've effectively done (very very crudely) is demonstrate that although you're working in LPCM the whole time, you've made a kind of DSD conversion (as i understand it anyhow... as i say above, i'm not a researcher) in so far as you've somewhat successfully traded off bit precision for sample rate and gone back without a fundamental change taking place (the waves are of course different - Audition/cooledit was certainly not designed for this).

it strikes me that, with correct dither, you're getting the best of both worlds with more bits - you can use the DSD concept (noise shaping, high sample rates) to "move" the noise floor out of the audible range, and you can use the LPCM concept for more convenient treatment of the signal (i don't know of any VST plugs or DAW progs that do DSD, at least at the consumer and semi-professional levels).

so really, to say that DSD is _better_ than LPCM when there's a demonstrable equivalence between the two, is somewhat strange. 1-bit noise-shaped LPCM at ~1.8mhz = SACD

feel free to pick at this argument all you want - i'm actually quite tired and perhaps didn't put this the most succint way i could. but definitely try the above technique, as i feel it demonstrates a fundamental similarity between LPCM and DSD, even if it's a qualitative one.

Linear PCM is crap, SACD is best

Reply #49 – 2004-04-04 20:18:27

This is truly one of the most amusing discussions I've read here in a while.

I cannot resist the need to reply to a couple of points dekkersj made in his arguments:

- Regarding your statement that PCM encoding cannot be transparent: To borrow a statement that gets tossed around occassionally around here, "Transparent" does not mean "perfect", it means "transparent". Ultimately, if no human ear is incapable of detecting sound differences below a certain level, any digital representation of an analog wave that can render it transparently for perception with such masking limits can mangle audio information below that level *however it wants* and no human could ever hear the difference--thus it would still be perfectly transparent even if it is not a prefect representation. A fully and completely transparent digital recording (sufficient bits and sampling rate) is perfect as far as the whole of humanity is concerned. Moroever, comparing LPCM's capabilities with an optimistic estimate of the human hearing capacity strongly suggests (if not outright shows) that it is such a completely transparent digital recording format, including the worst case effects of the quantization noise you say mangles quiet passages. Now, if Kang and Kodos were to land here tomorrow with much better ears than that of any human, we'd have to come up with something better to please our new alien masters.

- I don't see how you can reject sufficient dynamic range as a legitimate rebuke of your bit resolution concerns. You can row against the stream all you want there, but you're still going down the waterfall. You shoot down someone's argument that LPCM provides 144 dB of theoretical dynamic range by saying that a particular consumer DAC allows only about 117 dB of real dynamic range, yet you don't see that it is indeed here where your bit resolution arguments fall apart? Put it this way: how many bits would it take to encode a sound, of whatever frequency is most sensitive to the human ear, at 0 dB (threshhold of hearing)? For an undithered digital representation (which is the case where your arguments most apply), how would the quantization noise for that sound compare to the noise floor of the playback equipment? Would that noise *ever* be audible above the noise floor of that playback equipment? Looking at it another way: How many bits would it take to encode a 10 kHz tone, for which the human ear is much less sensitive and thus would have to be much louder than 0 dB to exceed the ear's threshhold of hearing? Would this make *any* difference in whether quantization noise can be heard above the noise floor of the playback equipment? I think it was you, in fact that made the following point in a related post, which makes me question the reasoning behind your arguments agains LPCM in the first place:

Quote

The established format (16/44k1) has a noise floor of 213 nV/sqrt(Hz) and this is quite high in comparison with 24/96 (576 pV/sqrt(Hz)) but still lower than a common audio amplifier.

It's important to note how the main technical deficiency of LPCM you're concerned with becomes so irrelevant in real-world usage, or non-existent when proper dithering is used, yet those of SACD are more of a problem because they cannot be entirely removed no matter what you do, and the system is more demanding in terms of signal processing and well-designed filters to sound well (leaving more room for error than LPCM does).

Notice