24/192 Downloads ... and why they make no sense?

[crisnee]

Teresa said "To my ears 24 bit has a wider, deeper soundstage, more ambiance and ease of presentation. If I put on a lossless 16 bit even from an audiophile label right after playing a 24 bit recording, I notice the music close-in, become congested with an added touch of stridency.

 

Now this is really weird, but several times recently I've had the exact opposite experience. In my head I described the sound of the 24 bit files exactly as you did the 16 bit files excepting possibly the touch of stridency. It was admittedly different music from what I'd been listening to at 16 bit (classical 24 bit downloads from the B&W music site), but it was so apparent that I started to wonder whether something was wrong with the 24 bit files--highly unlikely.

 

I can't think of any reason why this should have been so. I wouldn't have been surprised had they sounded about the same, but... In the end I attributed it to recording quality, but that still doesn't really sit well with me.

 

-Chris

 

 

 

 

 

 

[Julf]

Can we have more postings like this, please? Or does it violate the CA ban on facts?

 

[Miska]

Case 1) transient shorter than 22 µs - won't be encoded at all.

Wrong.

See above. Such a signal has spectral content below 22kHz. When the signal hits the ADC's anti-aliasing filter the part above 22kHz gets sliced off, and the part below 22kHz gets sampled.

 

No, it won't get encoded. Your Dirac-pulse of course contains all frequencies from zero to fs/2. You could calculate at which point it's energy would become so dysmal as to disappear into 16-bit noise.

 

Put in a band limited transient that doesn't alias, like a 5 µs square:

 

 

 

And when down-converted to 44.1 kHz sampling rate with proper anti-alias filter, output is straight line. Completely nothing.

 

 

[Miska]

Squares. This means that these transients should remain were we to perceive the "clearness" of the jangling keys. This is nothing about the "fineness" or refinement or "nice silk" sound, most people tend to think about, when talking HiRes. It is almost the other way around.

 

Hires allows you to reproduce squares better, because it can encode more of the odd harmonics needed to reproduce it.

 

For your enjoyment, I'll re-post this plot of 1 kHz square (posted earlier in another thread), encoded in pure DSD and played back through DSD DAC:

 

Harmonics disappear in the noise floor around 150 kHz as result of reconstruction filtering.

 

 

[Teresa]

"What system do you use to hear a difference between 24/44.1 and 16/44.1 recordings?"

 

NOTE: If you purchase a BIS 24/44.1kHz download from http://www.eclassical.com/ they include the 16/44.1kHz lossless version as well at no extra cost and you can compare for yourself.

 

Now to my audio system:

 

LOUDSPEAKERS

Infinity Reference Standard 7 Kappa, 12 inch woofer, 3 inch midrange, EMIT tweeters. Frequency Response 37Hz-45kHz +/-3dB

The 12 inch woofers give the weight and body to bass instruments that smaller cones cannot, in addition with 325 watts from my power amp into these wonderful speakers I have impact and power that is very close to concert hall realism. The power amp also has an extremely high damping factor (over 800) for excellent woofer control. The EMIT tweeter is perfect for high resolution sources and the midrange sounds sweet with the right recordings.

 

HEADPHONES

Sennheiser HD 580 headphones Frequency Response 12Hz-38kHz

Sennheiser HD 515 headphones Frequency Response 14Hz-26kHz

 

ELECTRONICS

AMC CVT 1030 Stereo Vacuum Tube preamplifier, Frequency Response with high level inputs 4Hz-80kHz - 3dB.

 

Adcom GFA-555II High Current amplifier, 325 watts per channel at 4 ohms, Frequency Response 1.7Hz-100kHz - 3dB, Damping Factor greater than 800

 

SOURCES

Yamaha Natural Sound DVD-S1700 DVD Audio/Video SA-CD player with Terra Firm Lite Clock mod - see my review http://www.positive-feedback.com/Issue46/terra_firma_lite.htm . Replaced the $1,700 retail tubed Xandak SCD-2 which was too problematic and down for repairs too often. The Yamaha's transport is dead quiet plays every single disc correctly every single time. None of my Sony's or the Xandak could do this. In addition in many areas the DVD-S1700 sonically surpasses the SCD-2, but the Xandak is still a little smoother due to it's tube output. All around the first SACD player I truly fell in love with.

 

MAC Mini computer for music downloads up to 24 Bit 192kHz.

 

Sanyo Flat Screen Color TV.

 

CABLES

Monster Cable Interlink 400MK II and M350i interconnects and Monster Cable Powerline 2 Plus speaker cables.

 

ACCESSORIES

AudioPrism CD Stoplight by Clear Image Audio™ Compact Disc Edge Treatment.

 

My listening room is 14 feet wide x 12 feet long with a ceiling height of 8 feet. My speakers are 8½ feet apart measured by their innermost point toed in slightly to get a more realistic image of an orchestral shell. Even with my speakers placed this far apart the phantom center image is solid and there is no hole in the middle. The speakers are almost two feet from the rear wall and I spent over six months moving them microscopic amounts at a time to find their optimal position and marked the spots. With the best source material the image is huge expanding beyond the outer boundaries of the speakers with excellent depth and height and ambiance in front of the performers filling up the rest of the room. Visitors have commented that my 2-channel stereo has the spaciousness that surround sound aims for but seldom achieves.

 

 

[PeterSt]

Digital depicts inifitely short time. Say 0.00000000µs.

 

Oh, it definitely doesn't, it represents a smooth transition from value to another of period 1/fs so that it doesn't exceed fs/2 bandwidth limit.

 

This looks like beating around the bush. I said "depicts" while you make that "represents". I said "digital" while you imply analogue.

I say : digital implies the worst, while a good method (subject here) is needed to don't let choke analogue.

The commonly used methods are not taken for granted by me. Why ? I see what happens in practice. I am (almost) sure you can do that too, but now you should really do it ?

(it is a somewhat difficult subject when certain amounts of post ringing are accepted, as are phase anomalies)

 

Sadly, it is the "representation" which is the whole subject here. You seem to be claiming that nothing in the world can be wrong with accepted methods within Theorems etc., while it is easy to see that not even a huge transient won't be found back unless in a general lump. Remember, a transient goes up only (per the subject in order).

 

Now I'm not sure what you think being a transient vs the signal seen by ADC.

 

What do you mean with "signal" now ? Ok, I suppose a "frequency". But in the end it is not important, if you'd only look in the resulting files. No function needed for anything. Just look. Btw, how to make that a function if it's only something which is observed ?

May you not understand what you see ? then maybe a disussion is needed why we see what we see. What *did* happen at the ADC concerned. Tough job.

 

 

 

[Miska]

What do you mean with "signal" now ? Ok, I suppose a "frequency".

 

What ever waveform that has been band-limited to fs/2 as required.

 

Which of course consists of mixed bag of different frequencies.

 

May you not understand what you see ?

 

See what? I don't read samples as line-connected waveforms because it's wrong representation. And samples are there only to represent coefficients against sinc function.

 

 

[Fokus]

"What system do you use to hear a difference between 24/44.1 and 16/44.1 recordings?"

 

With the system proven fine there are new pertinent questions:

 

1) how quiet is the room?

 

2) what is the usual listening level?

 

3) what is the exact provenance of the 16b and 24b versions,

i.e. who did the requantisation and what dither strategy was

followed?

 

 

[PeterSt]

... and not even sure from what. :-)

 

DSD is another animal, and I think we already agreed in advance that it will be able to do the job way better.

Not that this is not important of course, but I don't think the underlaying subject is - or should be DSD when it is all about transients and higher sample rates and such.

 

But do notice that DSD works out up to beyond the subject, only when the sample rate is high enough to not cause distortions. So, who is going to tell where this happens ?

Please let this be, and let's accept that DSD is okay for this all.

 

But not PCM. That is, not with the poor sample rates.

And as I tried to reason out (maybe I am not good at this at all) : 176.4 vs. 44.1 doesn't help much. Or not enough.

But I think we also agree upon 705.6 (768) being close to sufficient.

 

The majority of my personal subject is that it doesn't matter whether this was recorded "hires" or upsampled. Either do the same job, with now my suspicion that upsampling does an even better job than native Hires, if *only* the Hires is not sufficient to begin with. And since 176.4 is not ...

 

Additionally, with DSD the whole subject can't exist, because we *have* to accept 2.8MHz as being sufficient when we at the same time state that 768 is sufficient for PCM. I mean, DSD won't go lower ... (hence no upsampling can ever be in order to get it right).

 

That, in the mean time, the squares of your picture torture the speaker drivers is another matter.

Ah no, wait, this won't happen because of an analogue filter behind "the picture".

 

So, if anyone is able to tell what we are all talking about, I be happy to read about it;

Jangling keys can be represented nicely, but we better don't.

 

So yes, to open another Pandora's box. Hires sounds way too soft to me. There's no life in it. And don't say this is a subective thing. Tell me that I want to listen to distortion, and I will tell you (no one addressed in particular) that Hires is distortion. It is in my last posts (but merely implied) but it may become more clear from a picture like in Miska's post. So, NICE that all the transients or whatever we call them can be followed now.

Poor analogue.

 

Well, something like that.

Peter

 

[Miska]

Just been trying to read stuff on instruments with significant inharmonic components, e.g, percussion. How are these modeled / sampled / reconstructed?

 

It becomes band-limited representation of the same, so all the frequency components (plenty of those) below fs/2 will be encoded, while frequency components above fs/2 will be removed. This makes the transient slow down in terms of rise time.

 

The band-limiting also introduces ringing to the data. This ringing depends on properties of the anti-alias filter used. Steeper the filter, more it rings.

 

Like I've said multiple times, optimal case (IMO) is when the filter's entire ringing fit's into half-wave of highest frequency you want to encode. This way it won't ring at all within the frequencies of interest. For 20 kHz I can do it at 352.8/384 rates.

 

As PeterSt says, (inharmonic) attacks characteristic of music have a tremendously fast rise time, but the decay of the sound is not nearly as fast.

 

This is the reason I personally favor minimum-phase filters. These don't have any pre-ringing but fairly large and long post-ringing. Down side is that it produces phase shift as function of frequency.

 

Best approach is to use high enough sampling rate so that all original harmonics are encoded. There won't be band-limiting implications and thus the output is clean.

 

 

[PeterSt]

The 24 bit rate affords a longer word length and with 256 times the resolution of 16 bit this coupled with a higher sampling frequency that moves the actual sampling of the waveform further from the audible frequencies is what I feel accounts for the smoother sound of high resolution digital.

 

As long as you know that 24/88.2 has only two times the resolution of 16/44.1, that is fine.

24/176.4 is only 4 times.

Notice please that I incoportate acurracy here. It is determined by the lowest of the two "parameters", which is the sample rate.

 

[Miska]

only when the sample rate is high enough to not cause distortions. So, who is going to tell where this happens ?

 

Now the distortion products are all from analog electronics and below -100 dBFS (0.001%).

 

I mean, DSD won't go lower ... (hence no upsampling can ever be in order to get it right).

 

I've limited it to 2.0 MHz. But now I allow it to go up to 24.576 MHz (DSD512, with or without upsampling).

 

That, in the mean time, the squares of your picture torture the speaker drivers is another matter. Ah no, wait, this won't happen because of an analogue filter behind "the picture".

 

It would stress the tweeters for sure.

 

That plot is after all the analogue filters, so that's what would go to speakers.

 

 

[PeterSt]

No, it won't get encoded. Your Dirac-pulse of course contains all frequencies from zero to fs/2. You could calculate at which point it's energy would become so dysmal as to disappear into 16-bit noise.

 

Miska, as always, thanks for your supporting pictures.

But now I have a question;

 

While this now is about pulses and your remark seems valid, I personally don't think any long-distance (!) "transients" (bear with me !) will be eliminated by any filtering. So, assumed it is caught by the ADC, this implies some frequencies throughout. I think beyond the fs/2 limit.

Correct ?

 

Next, I don't see where the 16 bit noise occurs. That is, not in FFTs. This is similar to my remark (elsewhere I think) about the -141dBFS 16 bit signal just being there, which I obviously can't have seen when the 16 bit noise was there first. And no dither or anything was applied.

 

All I want to say is : we must be careful with theories vs practice, and I will be the last who says he knows better.

Btw, by pure coincidence the article in subject here "noticed" the same about this very much attenuated signal. And I didn't even start digging because I read the article (which I only did yesterday).

 

Peter

 

[PeterSt]

That plot is after all the analogue filters, so that's what would go to speakers.

 

I didn't dare to suggest that, but I hope you can see how very high frequency distortion is able to make the sound "softer". More sweet. This is why I asked Fokus to look what happens when these kind of pulses run into eachother. It will be one pile of wobbling mess, with sheer infinite repetetive cycle. So, not even a signature to the sound. But flattening all over the place *AND* doing that to the normal signal/sound in the mean time. This happens with speaker drivers (not rated at "squares") but in electronics also ...

 

[PeterSt]

Maybe some counterweight to the "sharpness" of DSD as output by the DAC (assumed it is able to follow properly) since it is my idea about it anyway :

 

DSD won't be as "hard" to squares as it may look (by all lurkers trying to grasp all the blah). This is because of the nature of DSD, always wobbling around the real (ever analogue) signal. So, this softens.

 

Don't ask me what happens when this softening element is seen as MHz "harmonics" and thus are nicely filtered out.

 

See ? I can make it so much more complicated -at least to myself- than, well, it really is ?

But beware when I am right ...

 

[Teresa]

"As long as you know that 24/88.2 has only two times the resolution of 16/44.1, that is fine."

 

16/88.2 has twice the resolution of 16/44.1.

 

With a 16 bit word, there are 65,536 possible values. With each single bit of greater resolution, the number of values double. By the time we get to 24 bit, we actually have 16,777,216 values. That is 256 times as many values.

 

That is why 24/44.1 sounds so much better than 16/44.1, so many parameters of music are improved by just increasing the bit depth.

 

[PeterSt]

It seems a bit strange that you apparently think to present me the math of these matters, as your only counterweight.

 

Remember that I mentioned accuracy rather explicitly. You may have missed that.

 

Regards,

Peter

 

[Julf]

"16/88.2 has twice the resolution of 16/44.1."

 

Only if you view it linearly, as your eye would. That's not how your ear works. One way of looking at tit would be to state that it adds one more octave of frequency range. 16/44.1 already covers 9 octaves from 20 Hz to 20 kHz. So you are increasing the frequency range from 9 to 10 octaves. So an increase of 11 %?

 

As to the amplitude resolution, an increase from 16 to 24 bits takes the dynamic range theoretically from 96 dB to 144 dB (in practice maybe 120 dB), so an increase of 50 %, not 25600 % as you claim.

 

And as PeterSt points out, you can't just multiply the gain in frequency range with the gain in dynamic range.

 

 

[Miska]

I didn't dare to suggest that, but I hope you can see how very high frequency distortion is able to make the sound "softer".

 

Well, I don't see any distortions on wideband spectrum analyzer plot. Above 150 kHz there's only noise floor...

 

I doesn't sound "softer" at all when there's real high frequency components present. This kind of square definitely doesn't sound soft.

 

When the sound IS soft, it can sound soft. When it is not, it doesn't sound like soft.

 

 

[Julf]

"the errors adds up. It accumulates through the chain. Even if your loudspeakers have 2% distortion, you are still able to distinguish between 2 different op-amps inside your DAC with distortion figures as low as 0,001%."

 

Sorry, I don't follow you. Even if we simplistically assume the accumulation is linearly additive, you would end up with 2.001% distortion. Are you saying you can hear the difference between 2.000% and 2.001% of distortion?

 

(In reality independent distortion doesn't accumulate linearly - it is actually a root-mean-square (RMS) function). So the real result is actually 2.0000002%)

 

 

[Fokus]

No, it won't get encoded. Your Dirac-pulse of course contains all frequencies from zero to fs/2. You could calculate at which point it's energy would become so dysmal as to disappear into 16-bit noise.

 

Is this dishonesty, incompetence, or were you not paying attention?

 

Read back. Where did I call for a Dirac? We were discussing impulses shorter than 22 us. And yes, the shorter the impulse the more insignificant its sub-22kHz spectrum becomes. I mentioned that too. Read back.

 

This does not detract from the simple fact that, thanks to the AI filter, impulses shorter than the sample period do get sampled.

 

 

Put in a band limited transient that doesn't alias, like a 5 µs square:

 

You show a square wave. Not an impulse, not a transient. Something periodic. It has a fundamental. The fundamental exceeds 44.1kHz.

It is bloody obvious that such doesn't get past the AI filter.

Again: dishonesty or incompetence?

 

 

Like I've said multiple times, optimal case (IMO) is when the filter's entire ringing fit's into half-wave of highest frequency you want to encode.

 

'IMO' indeed. What do you know about the audibility of the (pre-)ringing? There have been studies. In the midrange it is very very obvious. Towards higher frequencies it gets inaudible. Where is the crossover?

 

 

It will be one pile of wobbling mess, with sheer infinite repetetive cycle.

 

Please show us the wobbly mess and infinite repetition. Also please explain how this mess is perceived by humans.

 

So far we're only getting words from you. Very imprecise words, at that.

 

 

 

 

 

[Miska]

Sorry, I don't follow you. Even if we simplistically assume the accumulation is linearly additive, you would end up with 2.001% distortion. Are you saying you can hear the difference between 2.000% and 2.001% of distortion?

 

Of course distortion from transistors is more audible because it's mostly odd harmonics which sound bad by nature. While speakers tend to produce more even harmonics.

 

So the THD sums up, but from different types of components.

 

1% of even harmonics sounds probably less bad than 0.1% of odd harmonics.

 

 

[Julf]

"1% of even harmonics sounds probably less bad than 0.1% of odd harmonics."

 

Agree with that. But do you think you can hear 0.001 % of odd harmonics when added to 2 % of even (and odd, as not all speaker distortion is even) harmonics?

 

[PeterSt]

Again: dishonesty or incompetence?

 

Fokus, I understand that you are new here. But please don't pose it the way you do. No matter you think you are correct.

 

No, I am not a moderator here.

 

Thanks ...

Peter

 

[PeterSt]

Please show us the wobbly mess and infinite repetition. Also please explain how this mess is perceived by humans.

 

Why ? because you don't want to do it yourself ?

 

And better cut out the humans stuff - and studies for that matter. No study from yesterday is representative for this all.