i thought it is best to record around -18dbfs

[Reggie]

30db of boost?

If you make a graph where amplitude is the Y abd time is the X. Make the Y a logrythmic scale representing the possible voltages. (the positive side should be enough to prove the point) Now, superimpose an audio signal on top of it.

I call it "Cool Edit."
Not real sure what you are getting at here, but let me ask this: Below -60db, is +/-34 different possible levels enough to accommodate any possible analog waveform 0-22000Hz (-90db floor)? I hope your waveform happens to be in sync with the converter so that the wave just happens to be at one of the possible voltage levels at the exact moment it is getting sampled. Otherwise there will some rounding going on, getting more severe the lower in level. I always knew there was a % of quantizing error, but I guess I just never really knew where it came from or how it worked.

Another interesting fact on what those lower bit values actually are:
Value #0 = -96db
Value #1 = -90.3db
Value #2 = -84.28db
Value #3 = -80.76db
Value #4 = -78.26db

Weird, huh?


This book looks interesting in a nerdy way: http://books.google.com/books?id=VZ...H2ov&sig=3hPQQp7WZsdeKCWr5aw1TygrFXk#PPA37,M1
Check out pages 37-38, seriously. Seems to kind of say what I've been clumsily getting at.

 

[masteringhouse]

But I stilll don't get how the waveform can be reconstructed from inaccurate sampling that's biased either over or under. It would seem to me not that it would create a slightly inaccurate sine wave that can be dithered smooth, for example, but rather that it would - at least potentially - create an entirely different sine wave of different amplitude and slope altogether.

G.

Yeah I've had problems with this too. For example if you could sample at 40hz and there's a 20K sine wave (Nyquist) the sample can occur mid-cycle where there's a null creating no sound at all or somewhere between the max an min of the cycle creating amplitude less than it should be. The thing is audio in the real world is never clock accurate, so it all averages out to our ears (I think this is what Jay has been talking about). If it's consistently low though, it's quantization error and is inaccurate.

And now we're round circle. Why record hot? To avoid quantization errors that occur at low level signals. How important is this in the 24 bit world? Not so much if you keep the digital noise floor well below what analog can supply. Peaking at around -6 is definitely in this range of cushion.

 

[Reggie]

How important is this in the 24 bit world? Not so much if you keep the digital noise floor well below what analog can supply.

Word. I am just turning into even less of a fan of 16bit. Too bad there isn't 24bit Redbook CD-DA.
Then again, who cares!?

All very interesting though...

 

[Reggie]

Originally Posted by SouthSIDE Glen
But I stilll don't get how the waveform can be reconstructed from inaccurate sampling that's biased either over or under. It would seem to me not that it would create a slightly inaccurate sine wave that can be dithered smooth, for example, but rather that it would - at least potentially - create an entirely different sine wave of different amplitude and slope altogether.

G.

Funny you should mention this. I was screwing around with sample values and such the other day and I generated some pink noise within the DAW. I took note of the first 10 samples, dropped the gain by like 30db, raised the gain back up by 30db, and all the values were off by a surprising amount. So you can imagine what might happen to a waveform that isnt generated from within a digital application, being fed into such a system, values being rounded here and there.

 

[SouthSIDE Glen]

Is lookahead the answer?

I'm not done trying to find the answer to my own question and in trying to get to the bottom of Nyquist reconstructiuon, but I did come across a possible explanation. The explanation does kind of make sense, but I'm not convinced that it's the actual truth, as the source page contained other information that I thought did not sound quite right.

Anyway this source explained the reconstruction in terms of curve fitting (more on this in a moment), and said that the "averaging" that perhaps Jay referred to, and the "probability problem" that I thought I saw is taken care of by a method of lookahead on the samples. In other words, instead of looking at the samples serialy, it takes them in forward-looking batches and fits the waveform curve based upon the look-ahead batches of information, the extra data which would tend to reenforce or stremgthen the likelihood of fitting the right curves. To quote:

In reconstructing the original analog waveform (and also for manual curve fitting using French curve templates), the process can't take just one [sample] at a time but instead must size up several upcoming [samples], (looking ahead) at all times. For example we may work on samples 1-8, then work on samples 2-9, then 3-10, and so on. The closer any of the actual frequency content is to the sampling frequency, the more lookahead is needed (I don't know the formula for how much). Insufficient lookahead means that the reconstructed waveform might come out different from the original

.

This does make some good sense, and would explain why "amplitude resolution" is a false issue; curve fitting will result in the same curve regardless of the inaccuracies built into the individual samples.

However, I'm unsure that curve fitting even fits into the equation, and think that may be an erroneous explanaition. My limited understanding has been not that reconstruction involves curve fitting based upon fitting the sample amplitudes, but rather that the curves are "grown" based upon the addition of constituant wavelets extracted from the rough conversion in a Fourier-style filtering.

The search goes on...

G.

 

[Boogie Master]

Massive mastering


Is there a music site your not on? You must spend you whole night answering posts on the numerous ammounts of forums your on!

 

[snow lizard]

The idea that it can average out or smooth out the quantization errors correctly after only a few samples would seem to me to be dependant upon a balanced distribution of over-value and under-value in the quantization errors. For example, just to pull a number out of the air, if it took only four samples to guarantee good reconstruction, what happens when all four samples' err on the low side, which seems to me to be an entirely feasible event?

Another way to look at it would be to ignore the quantization errors entirely and consider that audio dynamics operating on less than 8 bits may have a tendancy to sound like an array of toothpicks in a cribbage board while somewhere over 20 it sounds more like a thousand chia pets on steroids.

I don't think I like either of those scenarios. Not a very good analogy perhaps. Might be something to consider if dynamics ever become popular again.


sl

 

[theblackBay]

lets See

i got this article from emusician.com

headroom over 16-bit—it was better to aim a little under 0 dB to help minimize the chance of going over, than to try to get the hottest possible level to disk.

“The old-school guys always say, ‘Go for the hottest level; use up all your bits,’ and it’s 6 dB per bit,” says Rogut. “My attitude is, if you’ve got 6 dB to play with, keep your level at least -6 dB or -5 dB.”

Rosado says he’s too often given vocal tracks to mix that are clipped, and he routinely tells clients to record their vocals with the peaks around –6 dB. “I prefer to leave some headroom on it,” he says. “Because if it’s too hot, it doesn’t give you room to play with.”

Addabbo and Stasium both said that they supplement the compressor by riding the mic preamp during recording. “I’ll always ride the vocal,” Stasium says. “I’ll push it up a wee bit, and then during loud passages, I’ll back it off. I want to get that level so that it sounds like a naturally flowing vocal.”





they say to record vocals kind of hot so u wont loose any quality. they suggested to keep the level around -5dbfs but everybody on here always says to keep the level around -18dbfs so isnt that too low?

Now I know I have commented on this forum just wee little too much of late but (and I’m not sure if someone has already said this)

but it's the same thing... or should I say logic tells me it is the same thing.

Exhibit 1.

In the above quote you will note the Producers/engineers say -6 db now I automatically assumed they were referring to -6db digital peak level yes?

Now the original poster has interpreted it to mean -6 db RMS no?

Now -6 -5 peak level digital is around -18db RMS so we are talking about the same thing.

and that concludes my speech.

I love the idea of riding the volume switch how avant-garde that was what real old compression was all about.

 

[theblackBay]

I'm just amused to think

that all this was inspired by those guys looking at the meter that they see on incoming tracking the Digital PEak level and you guys talking about a different one RMS.

But it's the same actual level..

But as they say it's not the destination it's the ride that counts.

 

[theblackBay]

I agree with -18RMS

“Everything is sweet at -18rms” as they say....

It even rhymes too!

ha ha .

 

[Farview]

Yeah I've had problems with this too. For example if you could sample at 40hz and there's a 20K sine wave (Nyquist) the sample can occur mid-cycle where there's a null creating no sound at all or somewhere between the max an min of the cycle creating amplitude less than it should be.

This is where oversampling comes in. The analog signal is being sampled in the Mhz range and is being paired down to the sample rate that is being stored.

I assume that there is a similar mechanism on the DA.

I wish I had more than a vague idea of how this all works. I used to know and be able to explain it, but the truth is that it doesn't matter.

 

[Farview]

Now -6 -5 peak level digital is around -18db RMS so we are talking about the same thing.

Only if the audio has a crest of 12db. Many things will fall into this catagory, like bass and vocals.

Some things that won't: Super distorted guitars, synth pads, sine waves, violins, etc.... These will hall have a much lower crest, maybe as low as 3db. If you make these peak at -6dbfs, the rms will be somewhere around -9.

Drums and other percussive instruments have very high crests, a snare drum peaking at -6 might have an rms somewhere around -30db. Percussive instruments are the only ones where the peak level is useful, because the bulk of the sound is the initial transient.

 

[theblackBay]

cool

in this case i tend to think the recording person (producer) was looking at the peak of the peak level indicator and seeing that as a rule it did not go over -5 or -6

the actual level average may have been much lower but the engineer did not care as long as it sounded good. so he says

"oh yeah -6 -5 whatever just don't peak it."

because if it were, as you correctly said the average, the level would be rather hot i.e -9 RMS the equivalent to ultra compressing then turning the make up right up, no curve solid wave.


i'm riding the same wave as you here the beach is way off and we are heading in the right direction. winds from the east.

 

[masteringhouse]

This is where oversampling comes in. The analog signal is being sampled in the Mhz range and is being paired down to the sample rate that is being stored.

Yep, this is kinda esplained in the previous link I gave in the section "DELTA-SIGMA MODULATION & NOISE SHAPING".

It's also one of the reasons why MEs may upsample from a lower sample rate before mastering.

 

[SouthSIDE Glen]

This is where oversampling comes in. The analog signal is being sampled in the Mhz range and is being paired down to the sample rate that is being stored.

I assume that there is a similar mechanism on the DA.

I spent some time up at my local Borders last night trying to research this a bit more. The only thing I found was in the EXCELLENT book that I've recommended before, namely "Understanding Audio" by Daniel M. Thompson.

As he describes the DA, he says that the "stairsteps" are first reconstruced from the digital data, then second, the stairsteps are oversampled, and then the new, higher resolution stairsteps are LPF'd, and this LPFing is where the final smoothed wave is made; i.e. the inplication that simply removing the high frequency artifacts from the oversampled stairstep is all they need to accurately re-create the wave. Not specifically said, but deeply implied in that passage, however, is that the reconstruction will include any original quantization or other error from the encoding; i.e. that any such "resolution" differences would survive, even if they are entirely dismissable.

That said though, Thompson did mention in passing at one point that there may be dithering injected on the A/D side that would wash the QE out. However, he did not mention that at all in the part where he actually described the A/D process in detail, so I'm not sure what to make of that. Dithering on the encoding side doesn't seem right to me.

I read that whole chapter on the whole A/D//D/A process two or three times over as I sat there and sipped my iced coffee, trying to connect the dots in my own understanding (ironic analogy intentional ), but I really have gotten only incrementally closer to an understanding of the Nyquist part of it and the whole amplitude with two samples question.

I did find a book on the Borders computer system that is specifically all about Nyquist theory and A/D//D/A conversion, but it was available on order only and cost something like $150 bucks. Yikes! And it's probably one of those where you get $150 worth of sinc equations longer than my arm and no actual explanation in English as to what it all means. I'll have to hit the library next.

G.

 

[masteringhouse]

G.

If you haven't already checked it out, see:

http://www.amazon.com/Principles-Digital-Audio-Ken-Pohlmann/dp/0070504695

 

[SouthSIDE Glen]

G.

If you haven't already checked it out, see:

http://www.amazon.com/Principles-Digital-Audio-Ken-Pohlmann/dp/0070504695

OK, a quick perusal online of that book seems to confirm the look-ahead idea. From page 26:

"In any case, the output signal is not reconstructed sample by sample, rather it is formed from the response of many samples" (and in a bow to what Jay said earlier) "in the same way that the ear does not hear each part of a waveform, but instead hears the average of the sound over time."

I think something just clicked in my head that, if true, I think would tie it all together in a neat bundle...

My missing piece to the understanding of Nyquist is that while only two samples per cycle are needed for the encoding, the decoding needs more than a single cycle - and therefore more than two samples in total - for that sample rate encoding to work. Which in real life is not an issue, because we're not dealing with single-cycle bursts.

Which is what we've been saying for a good page or two of this thread now, but - in my head, anyway - the idea the the two-sample per cycle thing applied to decoding as well is what keept tripping me up.

If this is true, than that would indeed wipe out "resolution" as any kind if issue whatsover in one fell swoop; any differences in encoding resolution would be eliminated by the averaged curve as filtered out in decoding.

As far as the balance of overs/unders, that would be mitigated by the fact that it's the frequencies closest to 20kHz that would be the most affected by any unabalce in overs/unders. The lower the frequency the more samples per cycle we have, and the tighter the curve fit. Even adding only a half-sample or one sample per cycle would increase the probability of a good curve fit signifigantly.

Does that sound about right?

G.

 

[Reggie]

G.

If you haven't already checked it out, see:

http://www.amazon.com/Principles-Digital-Audio-Ken-Pohlmann/dp/0070504695

Yeah, that's the one I linked a few posts back. Somehow you can pretty much read the entire book from that link it seems...
pretty cool stuff (for nerds)

 

[snow lizard]

OK, a quick perusal online of that book seems to confirm the look-ahead idea. From page 26:

I think something just clicked in my head that, if true, I think would tie it all together in a neat bundle...

My missing piece to the understanding of Nyquist is that while only two samples per cycle are needed for the encoding, the decoding needs more than a single cycle - and therefore more than two samples in total - for that sample rate encoding to work. Which in real life is not an issue, because we're not dealing with single-cycle bursts.

Which is what we've been saying for a good page or two of this thread now, but - in my head, anyway - the idea the the two-sample per cycle thing applied to decoding as well is what keept tripping me up.

If this is true, than that would indeed wipe out "resolution" as any kind if issue whatsover in one fell swoop; any differences in encoding resolution would be eliminated by the averaged curve as filtered out in decoding.

As far as the balance of overs/unders, that would be mitigated by the fact that it's the frequencies closest to 20kHz that would be the most affected by any unabalce in overs/unders. The lower the frequency the more samples per cycle we have, and the tighter the curve fit. Even adding only a half-sample or one sample per cycle would increase the probability of a good curve fit signifigantly.

Does that sound about right?

G.

It sounds good in theory. In practice, I think it would rely on the overall quality of the converters, the software and basically the whole thing as a system. Remember The Pipeline Effect.

The idea of the pulses or wave cycles averaging in amplitude makes sense. It's still speculation on a theoretically perfect PCM system.

PCM already seems to be more complicated than the 1 bit format. Now perhaps if the converter could come up with an interleaving algorithm or something that could calculate discrepencies on the fly and come up with an "over-over-under-over-under-under" strategy that could try to align the average of the wave pulses back to par in a listening enviornment, either we would be better off as a result or the system would bog down and develop other errors as a result of being designed more complicated than necessary. Obviously I'm not Dan Lavry, so I can't answer that one.

Now if someone goes and develops the DSD DAW, life might become a lot easier. I'm sure that if we ever get there, that system will have its own series of bugs to iron out. Until then, either we have 24 bit or we have to print tracks with an effective coping strategy.

Funny you should mention this. I was screwing around with sample values and such the other day and I generated some pink noise within the DAW. I took note of the first 10 samples, dropped the gain by like 30db, raised the gain back up by 30db, and all the values were off by a surprising amount. So you can imagine what might happen to a waveform that isnt generated from within a digital application, being fed into such a system, values being rounded here and there.

Seems kind of scary. It also illustrates that if you reduce a level in a DAW, you're introducing generation loss of a sort at that point. Coming back to the "multiple tracks sum" idea, if you print many tracks hot to counter quantization error only to have to attenuate them a lot later, your quantization errors should still be there except now they were generated after the fact instead of closer to the source coming from the converter. Now consider that clock jitter could lead to phase problems with the sampling process (perhaps a contributor to The Pipeline Effect?) and that digital will preserve any distortion that may have come from the preamp at the time. Any of these errors that average out will also sum in a busy mix, and probably have a better chance of being magnified when you attenuate in the box.

At that point, it might be useful to consider that if you're only recording something like an acoustic guitar and a vocal, you can get away with hotter levels and it might actually help unless your preamp starts screaming in pain when you print. Whereas if you're doing a full band, you might be better off to go very conservatively. Less distortion on the way in and less mangling after the fact. It would make sense to want to avoid having to boost levels in particular when mixing, but also to try to minimize level changes in either direction.


sl

 

[SouthSIDE Glen]

PCM already seems to be more complicated than the 1 bit format. Now perhaps if the converter could come up with an interleaving algorithm or something that could calculate discrepencies on the fly and come up with an "over-over-under-over-under-under" strategy that could try to align the average of the wave pulses back to par in a listening enviornment...

I had several things to say about this, but like jr, my brain simply hurts now after beating it agaist the Nyquist limit for the past 48 hours. This is like quantum physics, there are five people on the planet who actually understand it, and four of them are lying.

The sad irony is just when I have PCM and Nyquist finally fully understood, it'll be obsoleted by 1-bit anyway. It'll be like someone just now figuring out RIAA equalization curves .

G.