DVD: Tomorrow's VHS, or Betamax?

I’d point out that the reason many people hear distortion on CD reproductions of things like female vocals and cymbol crashes has more to do with the increased dynamic range of the CD coupled with poor electronics and speakers. A CD puts some heavy demands on the other components in a system, which may not be up to it. In the case of a cymbol crash, the slew rates required of the amplifiers to accurately reproduce it are very high, and you may just wind up with clipping instead, leading to the distorted sound. Likewise, when you have a dynamic range of greater than 90 DB, not many people realize that even normal volume levels can require peak power above the rated power of the amp, resulting in more clipping.

Also, there seems to be a tendency on new recordings to record at near-saturation levels. Listen to Alanis Morrissette’s ‘Jagged Little Pill’. I’m tempted to call the album a faulty mix because it doesn’t seem much compression was done, leaving a very wide dynamic range. It gives it a unique sound, but makes it very hard to reproduce without driving equipment into clipping.

Moonshine ponders thusly:

Audio devices, such as say, an amplifier, or the electronics inside a CD player, are not perfect devices. There is a certain limiting ratio between the strongest and the weakest signals they can represent. If a signal is too weak, it just will get lost in the internal noise of the device. Think of it like whispering to somebody 50 feet away when there’s a jackhammer running right next to them - your signal (the whisper) has been lost in the noise (the jackhammer). If you raise your signal level, say by shouting, then you might be heard.

Similarly, there is only so much signal to noise ratio available in audio electronics. Even under perfect conditions, if you eliminate all stray RF noise, there is some noise left, from random electron electron fluctuations in electronic components, and this is a function of temperature. It’s called thermal noise.

Given how much noise is present from this source, one can predict how many bits are necessary to quantize a signal to this accuracy or better at that temperature. This is around 18-20 bits for good quality electronics. Sampling with, say, 32 bits, is massive overkill if you are limited in other parts of the system to a much lesser S/N ratio.


peas on earth

Chef Troy wrote:

In 1995, the first Dolby Digital laserdiscs started coming out. (It was called AC-3 at the time, but it’s the same thing.) They accomplished this by usurping one of the 2 old analog audio channels (!) that’ve been part of the laserdisc format since its inception.

The remaining analog channel is used either for a mono mix-down of the stereo channels so that really old laserdisc players without digital audio can still play it, or for a “running commentary” by the director or producer as part of a “special edition” release.


Quick-N-Dirty Aviation: Trading altitude for airspeed since 1992.

bantmof, thanks for the lesson. I knew all the physics you discussed, and most of the technobabble about audio, but never really pieced the two together. I learned something bringing me closer to living up to my username. I’m gonna try and ask a couple of more questions.

What exactly do the bit/channel rating mean. I assume its not comparable to a bit when discussing computer infomation.

Is there a number rating for the equipment which describes the accuracy of the sound reproduction that you have so nicely explained.

We’ve got Watts/channel, bits, bits/channel, ohms, Nx’s oversampling, samples/second.

Weave thses puppies together for us.

CD players use pulse-code modulation to convert analog data as digital. Basically, what happens is that a sample of the input signal is taken, and then instantaneous voltage level is converted to a binary number and stored. Then another sample is taken, and that value is stored, etc.

The number of bits in the sample determined how many discrete levels you can distinguish. If you use 8 bits, you can only have 256 different levels. If you only had 2 bits, you could have four different levels. So the fewer the number of bits, the less accuracy. You wind up with quantization error.

Here’s a diagram to ilustrate the digitization of a wave using PCM:
<center>
<img src=“http://www.rasip.fer.hr/research/compress/algorithms/fund/pcm/pcm_1.gif” border=0>
</center>

Here’s the wave that can be re-created from the quantized data:

<center>
<img src=“http://www.rasip.fer.hr/research/compress/algorithms/fund/pcm/pcm_2.gif” border=0>
</center>

Okay, notice that the slope of the original signal changed between time 1 and 2. Since there was no sample taken between those two points, that information is lost. If you look at the bottom image, the curve is uniform between 1 and 2. That’s because the sampling rate was too low.

Now look at sample #6. Because this signal is encoded with only 4 bits, any signal levels that don’t match the level represented by one of the bits has to be rounded up or down. In this case, it’s rounded down, so the output wave has a lower amplitude at that peak.

If we used only 2 bits to encode the signal, we could only represent four different levels. If you look at the diagram and imagine having to round up or down all the values that fall on 1,3,5, you can see that the output wave would look a LOT different than the input.

Now I hope all my HTML comes out correctly.

I think I see.

First let me point out what seems to be an error. If 2 bits has 4 levels (2[sup]2[/sup]), and 8 bits has 256 levels (2[sup]8[/sup]), wouldn’t the diagram show a 3 bit system? (Granted I’ve never seen a bit count thats not a multiple of 4) 2[sup]3[/sup]=8?

If this presumption is correct, and the chart shows a time interval in seconds. Then the graph shows a 3 bits/channel at 1 sample per second.

Am I close?

PS I knew how a signal is digitized, but I’m just trying to decipher the crap written n the side of the CD/DVD players box.

Yeah, it’s a 3-bit system. Sorry about that.

Your understanding of Fourier stuff is much beyond what I know, bantmof, but I did manage to follow. :slight_smile:

I keep thinking something is wrong with what you say, mostly because I think I’m right. This, of course, is a really stupid way to defend a theory, but here goes…

I agree with this statement, however, I think there’s more to it: not only does a device have to be able to produce the sine wave at the highest frequency, but also be able to produce sine waves at all frequencies below that frequency. This is the crux of my argument (at least in this post).

The extreme case being, say a signal generator with a variable amplitude that can create a 30KHz sine wave–this does not imply it can more accurately reproduce waveforms than an 8-track tape player which rolls off around 8KHz.

Likewise, if you had 100 or 1,000 signal generators, each of which generated a particular frequency sine wave (I’m assuming they’re evenly spaced across the 20Hz-20KHz range) you could produce most waveforms pretty accurately, but there would be acoustical losses you could hear that you couldn’t see on an oscilloscope.

Both CD’s and record players are limited to about the same lower limit of 15Hz or so. They’re tied there.

Aside from mechanical and electrical harmonics in a record player, I think you’ll agree it can reproduce any sine wave accurately between 20Hz and, say 15KHz (i.e. record a sine wave on a record then play it back). Above 15KHz, the response rolls off at some odd curve (that is, not a pretty 3dB/octave or something like that), and it would probably be impossible to reproduce any measurable wave above about 40KHz (which is another argument altogether–CD players have a hard limit of 22KHz while record players roll off up to a higher frequency…maybe another day.)

Consider my earlier example about the 11KHz wave ideally sampled at 44K-samples/second–the normalized output was a repeating function {0,1,0,-1}. This works great and reproduces a perfect sine wave.
Now consider an 11,400Hz wave. (I had to cheat and plot it out, below.) Notice how the wave generated from the 44k-samples/second has some additional frequencies in it. These additional frequencies tend to cause noticeable distortion.

<IMG SRC=“http://www.servtech.com/~jayce/images/CDwave.gif” WIDTH=“400” HEIGHT=“270”>

The problem here is that the CD player can not generate any arbitrary sine wave between 20Hz and 20KHz. Rather, it is only able to simulate certain frequencies.

My theory is that to generate an arbitrary waveform based on a discrete Fourier Transform, the equipment used to reproduce the waveform must be able to generate any sinusoidal wave up to an arbitrary cutoff frequency. I kind-of showed that a CD player is not able to generate an arbitrary sine wave within its designed frequency range, and kind-of showed the record player is. I conclude that the record player can more accurately reproduce an arbitrary waveform.

Basically what I’m saying is a CD player has more high frequency distortion than what a specifications chart would lead you to believe.

Ok, now go ahead and poke holes in my already not-airtight argument. :slight_smile:


Hey, aren’t you supposed to be at work?

Yep, that’s true - it’s possible that some devices can have “holes” in their bandwidth. CD players (to a pretty good approximation) don’t however - they have a flatter bandwidth response than records by far.

The electronic parts of the CD player (the D/A’s, etc) might be limited to 15 Hz, but the digital audio itself on the CD is good down to DC.

Also, record players have a lot of issues to contend with below about 60-80 Hz or so. It doesn’t make it impossible for them to reproduce lower freqs, but it makes it trickier.

In the real world, I think CD’s win and win big on both the high and low ends of the spectrum. In optimal lab conditions, a record player probably wins at the high end, being able to have some level of reproduction over 22,000 Hz, where a CD has a hard limit there.

If I understand the intent of your graphic here, you’re talking about the fact that we have discrete samples at regular time intervals, right? It turns out that sampling (by itself) does not result in the loss of information for band-limited signals, as long as you sample at >= 2X the highest frequency present in the signal. Let’s first talk about a theoretically perfect world, and then the real world.

In our theoretically perfect world, we sample with perfect precision. I.e, we are not limited to the 16 bits that our CD player is, we have prefect filters, and perfect D/A’s. In this case, if we band-limit our input signal to less than 22 KHz, we can perfectly recreate it even though we’ve sampled it at discrete intervals and “missed” some of the peaks of the waveform where they didn’t line up with our sampling points. One way to think about this is that we’re going to fit a curve to our samples, not just blindly draw lines between the sample points - because we know the limiting frequency of the original signal, we know how to “fill in” the signal between sample points. We’re not going to just “clip off” the tops of curves and so forth as the dotted lines in your graph imply, because we know that would generate higher frequencies than were in the original signal. It’s possible to perfectly reproduce, say, a 11561.34663 Hz sine wave, even with discrete sampling. One can prove this mathematically, which is what the Nyquist theorum is all about - I might even be able to do it if you give me a while to dig out some old lecture notes.

Now let’s look at the real world. In real life what probably happens is that the D/A does generate something close to your dotted lines, missing the peaks as you show, and we get a signal with high freq components. But then we run this signal through a “brick wall” low pass filter and viola, as if by magic, our peaks come back in the right places, because we’ve removed the high freqs that were clipping them off like that. We don’t have perfect brick wall filters, but overall it works pretty well and what we get out the end much more closely resembles the orange lines than the dotted ones in your graph.

Ther are other sources of error though. Not only don’t we have perfect low-pass filters, we don’t have perfectly linear D/A’s, and we quantize at 16 bits, thus incurring amplitude error. I.e, maybe the value we measured as a discrete amplitude of 15621 was really supposed to be 15621.49. Obviously the magnitude of this error as a percentage of the signal increases as the signal gets weaker.

One can then ask what percentage of the reconstructed signal is error. Leaving aside some techniques like dithering, the maximum ratio at 16 bits is about 98 dB, and of course decreases (i.e, the error ratio increases) as the input signal becomes quieter. (I.e, an error of 0.5 out of say 28000 isn’t much, percentage wise, but the same error of 0.5 out of 3 is rather a lot!)

Of course, records also suffer from error which limits their S/N ratio, but they fare far, far worse then CDs do. Just from memory, I recall that ratios of 50 to 60 dB were considered good for records. CD’s win bigtime when measuring this sort of error. If you reduced your CD down to 9 or 10 bits instead of 16, it would have a more record-like S/N.

So I guess the summary here is: (1) The imperfection in the reproduced signal doesn’t result from the discrete sampling, as you imply. There are some gains to be had from oversampling, but at least in theory, and to a good approximation even in practice, increasing the sampling rate does not get you a more accurate reproduction of signals under 1/2 the sampling rate. And, (2) increasing the level of quantization does help both in theory and practice, at least to a point. But even 16 bit quantization is very, very good, and is almost never a limiting factor. And also, (3) it’s quite possible for a record to sound much better than a CD, but in all likelyhood this isn’t because the record is reproducing the signal more accurately. It’s much more likely because of some feature of the signal itself (perhaps the way it was processed beforehand), or perhaps because some types of errors induced by the record are pleasing to the ear (this can happen!) There are a zillion other potential factors that might make some particular record sound better than some particular CD.

Always glad to oblige :slight_smile:

peas on earth

Hopefully this is as amusing to everyone else as it is to us. I am a bit wearied, but I contend that the above-8KHz data sounds better on a record than on a CD, and by gum, I’m going to figure out why! :slight_smile:

Realize, of course, that I am only nitpicking on the high frequency reproduction of CD’s vs. records. CD’s win in basically all other ways (durability, S/N ratio, dynamic range, low frequency reproduction, ability to be duplicated, etc.)

Ok, on to the meat…you concede that:

I’d wager that this is also true in non-optimal non-lab conditions. What is important here is that even though the frequency response of a record rolls off over 20KHz (and probably starts at around 12KHz to 15KHz for records after they’ve been played a few times), it can indeed reproduce frequencies above 20KHz. CD’s, however, must be limited to 22KHz, both because Nyquist guarantees it and it sounds like crap if you don’t do very steep filtering (thank you DSP and your glorious 20dB+/octave rolloff).

This is important because an arbitrary waveform with a base frequency of less than 20KHz has, in all likelyhood, components (ala DFT) far higher than 20KHz. When the hard limit is imposed in a CD, these components are lost. A record player however, can reproduce them–not perfectly or with the amplitude of the original signal, but at least in part.

Likewise, the human ear with its little bones and stuff tends to have a frequency response which rolls off around 20KHz (usually lower) but nonetheless can pass the higher frequency data to the inner ear.

Don’t believe me? Listen to a signal generator. Listen to, say a 15KHz or 16KHz wave and change the waveform. Even at those frequencies you can distinctly hear the difference–hence the mention of “s,” “sh,” and cymbal hits sounding alike on CD’s.

Now, I have to disagree with you on the magic of filters restoring the waveform I plotted. First I thought, “dang, you’ve got a point there,” then I realized that the signal presented is a 11,400Hz waveform (carefully chosen by educated guesses and trial-and-error to best demonstrate the distortion). A 22KHz brick wall filter would not affect the 11,400Hz waveform very much.
Even the slope around the 0.2ms mark represents a component of less than 22KHz, except maybe for the sharp turn at the sample right after 0.2ms. Nonetheless, the CD player would reproduce the signal similar to the dotted line. What I was getting at was that frequencies whose harmonics are near the sampling frequency are more distorted than others. I haven’t tried it (signal generator + computer with sound recording…) but I bet you could record a signal around 11KHz which would noticeably warble when recorded digitally.

In the end, the point is kind of moot. From about 1984 on, almost all music passed through some digital phase, and now, almost every band records directly onto DAT (48K-samples/second for some reason). When this stuff is rendered onto record, the record accurately reproduces all the digitally-induced high frequency distortion. :slight_smile:

My A-B tests for friends are limited to older stuff, and often it’s a duel between Paul Simon’s Graceland on CD and record. I still say the record wins–except for minor things like hiss, pops, rumble, skips, fragility, dynamic range, sensitivity to physical vibration, …

Hey, aren’t you supposed to be at work?

Well… it does. Given arbitrary quantization and lowpass filters on both ends, you can reproduce the signal perfectly. This is something that you can prove mathematically, as any engineering major (myself included) has had to do to pass his digital signals classes :-). It’s not really a matter of opinion. The real world is imperfect of course, but you seem to be denying some very basic DSP theory - stuff which is well understood in the engineering world.

It does affect the reconstructed result from the sampled data, because that waveform has many higher frequency components. When you remove them, you get the original signal back.

Would it help if I posted the proof? I can’t do it off the top of my head any more (too many beers since college :-), but I can probably re-learn it in just a few minutes if I can find some old lecture notes. It’s a proof that shows you can perfectly reconstruct your 14400 Hz sine wave (or any arbitrary bandlimited signal) perfectly even with time-discrete sample points. You’ll need to know convolution operators and a little calculus to follow it.

[quote]
I’d wager that this is also true in non-optimal non-lab conditions. What is important here is that even though the frequency response of a record rolls off over 20KHz (and probably starts at around 12KHz to 15KHz for records after they’ve been played a few times), it can indeed reproduce frequencies above 20KHz. CD’s, however, must be limited to 22KHz, both because Nyquist guarantees it and it sounds like crap if you don’t do very steep filtering (thank you DSP and your glorious 20dB+/octave rolloff).

[quote]

Most records were created with a signal band-filtered to about 15 KHz before production, to save wear and tear on the cutting head, so for the most part these higher frequencies were lost before they ever got to the vinyl. And at high frequencies you can hear, over 10 or 12 KHz but under 20, records have truely frightening levels of distortion at any sort of signal level at all. There are a lot of physical reasons for this, but that’s probably left for another thread. But the point is it’s been measured, and CD’s reproduce high frequencies more accurately than records, not less.

Another useful thing to note is that a few years ago, some company (JVC maybe?) did a test where they recorded 16-bit CD quality digital audio from a high speed reel-2-reel master of a concert. In double blind tests, people could not tell the difference between the 16 bit audio and the analog master.

I guess maybe what you are doing here is hearing some records that sounds better than the same music on some CD, and looking for reasons to explain the CD’s “inferiority”. But the technical reasons you’re coming up with don’t hold water. I’ve been way deep into this stuff (like we’re at 30,000 ft and once upon a time I understood it to about 5 feet) and what you’re claiming about sampling just ain’t so.

But here’s the thing! Look at it this way - there might be other perfectly plausable reasons why you might like a bunch of recordings better on records than on CD. There’s nothing wrong with that! We don’t need to invent invalid reasons, when there are enough valid ones to go around. I’ll even help you think of some :slight_smile:

Oh, if anybody else was reading I suspect we bored them into a coma by now :-). Anyway, I’ve enjoyed the debate; hope you have too.

peas on earth

Aw crap, I sort of blew the quotes there. Anyway, most of the stuff in that huge quote block was my text, not quoted.

Also, here’s a web page I found with a nice, non-technical (i.e, no math whatsoever) description of this issue:
http://www.bores.com/courses/intro/basics/1_antia.htm
Two relavent quotes from this page:

And,

I think that’s addresses the point contained in your graphic.


peas on earth

My only remaining argument is that it’s the higher frequency response of records that allows “arbitrary waveforms” to be reconstructed accurately at high frequencies. I don’t know anything about how records are made (especially, how the audio data is processed before being applied to the surface). Then again, it could be that I’ve just got a crappy older CD player which doesn’t have a good brick wall filter.

I kick myself many times on your behalf because the plot I made is really stupid. The ouptut from the D-A converter would look more like a stairstep function, and when that signal is filtered, it’s more obvious how to get the original waveform back. The plot looks more like:

<IMG SRC=“http://www.servtech.com/~jayce/images/CDWave2.gif” HEIGHT=270 WIDTH=400>

Note that I didn’t do much correction on this and the digital samples lag the input signal by half the time of one sample.

Regardless, I still think pulse-width modulation would be a better way to record audio…it’s simpler, and due to the high frequency of the pulses, you can use cheaper filters to remove the high-frequency data added by the technique. Similarly, higher sampling rates would make me happier too for similar reasons.

I always try to remember that theories may not be applicable in all cases. In the theoretical world, DFT works perfectly. In the real world, it is a useful tool. When it’s applied to human perception, the theory may not hold up, or more likely, underlying assumptions leading you to use the theory may be incorrect.

In this case, I disagree with the assumption that human hearing is limited to 20Hz-20KHz. Conscious perception of a tone may be limited to this range, but there are probably other factors at work. The ability to distinguish waveforms near the upper limit of hearing (i.e. 18KHz tones) leads me to believe, based in part on DFT, that human hearing has a higher bandwidth than the commonly accepted 20KHz limit. Oops…I’m starting to make another topic. :slight_smile:


Hey, aren’t you supposed to be at work?