Ok math nerds....

In the meantime....I'm over here pressing the red button every time I hear something that SOUNDS good.

I don't care WHY......just that it does.

I think it is both the math and the mixing.

The math's in the filtering. As we know, an AD converter needs to anti-alias filter in the analog domain before conversion. And if you look at a filter, it's cutoff frequency is defined by its -3db point. Meaning if you want to reduce alias frequencies below the level of your softest dynamic, lets say -100db, you have to start filtering a little before the Nyquest limit. On the order of 1.5 to 2 octaves, depending on how steep (ie, expensive) the converter filters are. The steepest filters I've ever seen in an AD converter having 8-pole 48db/octave filters with most commercial high-end ones using 5-pole 30 db/oct. So with a 44.1k sample rate and a 22.5k Nyquest limit, you'd be hard pressed to get an alias fee full-level response at 11K. And don't forget about the other little nasty mathematical characteristic of filters... phase shift in the pass-band as you approach cutoff. And if you try to steepen cutoff with a little overshoot as is common with most ADC's (eg, Chebyshev response), you get ripple in the passband. Problems which move conveniently above the audible range if you simply double your sampling rate. And as an added bonus, higher frequency anti-alias filters are easier to design and cheaper to build. Mathematically, its a win-win situation.

Mixing enters into it because as you add tracks together which each have the above problems, particularly the phase-shift and ripple issues, the destructive interference adds up causing not only high frequency cancellation, but a general collapse of imaging and focus... most commonly perceived as the stereo image turning into mush. I think Sonar's 64-bit mixing bus helps a little with 44.1k/24 tracks by limiting the error propagation, but its absolutely golden at higher sample rates.

The interesting thing is that if you record and mix at higher sample rates, and then downsample to burn to CD, much of your winnings in the way of frequency response, clarity, and imaging make it out the speakers on playback. And you're poised to move your material to DVD which might have better converters supporting high res playback.

The best analysis against hi-res audio that I've seen is by Dan Lavry who many believe makes the best ADC's on the planet... but he's arguing against 192k sampling in favor of a 96k upper limit...
http://www.lavryengineering.com/documents/Sampling_Theory.pdf

Howard

256x oversampling is what's usually implemented in cheaper converters which almost universally use a 20-bit internal resolution. The main object is to increase the resolution by averaging the extra samples. 256x yields an extra 4 bits for a 24-bit output. Works well in theory as long as the noise spectra is constant. Which it never is. Full resolution (24-bit) oversampling could accomplish a preservation of frequency response, however. I think some of the premium discrete converters do that. But it does introduce new problems. Reminds me of the 3 laws as I originally learned them... you can't get something for nothing. The best you can do is break even. And you never will.

Which is kind of the main object of Lavry's analysis. That oversampling and its decimation filters compromise converter accuracy the faster they go. He contends you lose conversion accuracy when you oversample. And talks about a fundamental trade-off between accuracy and internal processing speed. Which translates to a frequency response trade-up if that was the object of the oversampling in the first place. But if the object was to cheap-out with a internal 20-bit design, they'd cash in the frequency response preservation for 24-bit output width. And lose accuracy to boot. Welcome to the worst of all worlds.

The internal design of the TI PCM4222 is not clear on its data sheet... I suspect its the usual 20 bits because its cheap and the data sheet only mentions 24-bit when describing output specs. If so, its 128x oversampling would be sub-optimal.

Howard

Hi, AndyW. I had to dig into that a little myself because its a technique not used by my converter which is a 10 year-old Apogee PSX100. And not used by others I've considered upgrading to like the Benchmark or Lavry. I think its because these are premium true 24-bit converters that wouldn't need to make up bit-depth resolution with extra samples. But they do run at 2x or 4x at full 24-bit resolution and I think they just downsample to yield a fuller frequency-response 44.1k/24 sample.

Near as I can tell all the 64x, 128x, and 256x oversamplers use 20-bit converters. I think what they do is take a series of snapshots of the signal through their 20-bit window, dropping their numeric voltage reference by 1 binary unit for each successive snapshot. So 256 of these samples being 2**4 gives 4 more bits of low-voltage resolution. Problem is that when you're done you still only got 1 snapshot of your upper-voltage picture so you can't possibly get full relief for the analog anti-alias filter situation. Though I assume the digital decimation filter that knocks them down to 1 composite sample helps a little, in a fudged kind of way. Guess those extra 4 clear bits and the frequency response they rode in on are what the gear slutz prize because they all seem to shell out for the real deal.

The other part of the equation is the mixing. And I can tell you that even with my Apogee whose 44.1k/24-bit performance probably blows away any flim-flam 20-bit converter's, listeners seem to have no trouble picking out high-res alternate mixes.

Howard

ORIGINAL: hv

256x oversampling is what's usually implemented in cheaper converters which almost universally use a 20-bit internal resolution. The main object is to increase the resolution by averaging the extra samples. 256x yields an extra 4 bits for a 24-bit output. Works well in theory as long as the noise spectra is constant. Which it never is. Full resolution (24-bit) oversampling could accomplish a preservation of frequency response, however. I think some of the premium discrete converters do that. But it does introduce new problems. Reminds me of the 3 laws as I originally learned them... you can't get something for nothing. The best you can do is break even. And you never will.

Howard,

I'm not following you. If all the converters oversample at 64 x fs or more, why did you mention all that you mentioned in the previous post about analogue anti-aliasing filters?

128 x fs seems to be used for the top converters of all the major parts manufacturers. (Cirrus Logic CS5381, AKM ak5394A and the TI mentioned previously)

I don't know of any converter that does 128 fs (or 256 fs) at 24 bit. It doesn't make sense to me. You get all the linearity issues of having 24 bits and little or no benefits AFAIK.

Which is kind of the main object of Lavry's analysis. That oversampling and its decimation filters compromise converter accuracy the faster they go. He contends you lose conversion accuracy when you oversample. And talks about a fundamental trade-off between accuracy and internal processing speed. Which translates to a frequency response trade-up if that was the object of the oversampling in the first place.

Are you suggesting that Lavry's designs sample at the base rates?

The internal design of the TI PCM4222 is not clear on its data sheet... I suspect its the usual 20 bits because its cheap and the data sheet only mentions 24-bit when describing output specs. If so, its 128x oversampling would be sub-optimal.

It uses a 6 bit modulator. This is typical. This chip only costs about $15 but that doesn't make it cheap (relatively speaking) or sub-optimal.

The other part of the equation is the mixing. And I can tell you that even with my Apogee whose 44.1k/24-bit performance probably blows away any flim-flam 20-bit converter's, listeners seem to have no trouble picking out high-res alternate mixes.

The PSX 100 uses a Cirrus Logic CS5361 ADC chip. This doesn't sound as good as the TI PCM4222 (or the newer CS 5381).

I don't know why you are talking about flim-flam converters. Nearly all the quality audio interfaces/converters use chips from the three companies I've mentioned. There is nothing flim-flam about these chips.

The Benchmark ADC1, if you were referring to that in your last post, uses the AKM5394 chip.

UnderTow

ORIGINAL: hv

My impression is that it isn't so much what chips are used but how they're used in a converter's design. My PSX100 and the Lavry designs run at 2x (96k max)

Some of the Lavry's can actually run at 192Khz. It just isn't documented and there is no front panel indicator. The PSX doesn't (as far as I know) but it is an older design.

and document their analog anti-alias filters at around .56 that rate and yield a full 24-bit sample at their top rate.

I believe that for the PSX 100 that the filter is in the converter chip. I don't know enough about the internals of the Lavry converters to know exactly what Mr Lavry uses. (The Blue and Black converters are chip designs as far as I know. Not discrete).

I made a mistake in the previous post. It is the Rosetta that uses the CS5361 chip. (An older design chip that has been superseded!)

For the Rosetta, the 24 bits is the output rate of the CS5361 chip. It is still a multi-bit (not 24) oversample design. The extra bits are achieved in the decimation process. (I believe the PSX 100 might use the same chip but it might use an even earlier generation chip like the CS5343).

The Benchmark runs at 4x (192k max) as do Apogee's newer converters like the Rosetta192. Among these premium converters, the gear slutz folks rate the Lavry tops in sound followed by the Benchmark and the PSX100.

Again, the Benchmark runs at 128 fs until decimation. I'm not sure about the Lavry's or the PSX 100 but I suspect something similar.

And interestingly rate the Apogee 192 converters below the PSX100.

That might be explained by a radical new design. I don't know.

Don't get me wrong, though. I'm not sure I agree with all of Lavry's reasoning since he doesn't seem to take into account the high res mixing factors which might outweigh other considerations as the track-count climbs. But I have done other folks' mixdowns of tracks recorded with Lavry and Benchmark converters and there's no doubt in my mind that the Lavry is indeed tops. So I think he must be on to something.

The discrete Lavry Golds or the chip based Black and Blue designs that can run at 192Khz? Keep that in mind as it is very important to determine what is causing the improved sound quality. Also, increasing the sampling bandwidth is not an increase in resolution! That is just marketing hype.

The reason I think 128x oversampling in a 20-bit converter chip is sub-optimal is because its technically insufficient to fill out 24-bits. It'll only get you 22. This is probably why Lavry converters have an led that illuminates when it detects 22-bit audio data hiding in the 24-bit format which you might want to treat with a little dither.

But the chips I mention are not 20 bit designs. They are multi-bit oversampled modulator designs that deliver 24 bits output. Most of them are better than the chips used in the some of the converters you mention.

Anyway, the ADC chips used is only a (small) part of the entire design. The best converters achieve their quality and sound because special attention has been put into the design of the entire system. (Analogue stages, clock, power supplies, shielding, earthing etc).

UnderTow