Gardo:
If the problems with PCM are easily solved, why haven't they been? That's a real question. If I understand DP, he's saying that removing aliasing effects in the 20khz-24.1khz band takes care of digital "glare." How would this be accomplished? Better brick-wall filters? Aren't those aggressive digital filters part of the problem?
BP:
In order to "solve" the PCM problems, it is required that a certain amount of care is given to all steps in the production chain. It is currently no problem to build a set of AD/DA converters that avoid all of the pitfalls in PCM. However, until everyone takes the same actions to get their digital filters right, these issues can reappear at any time.
Here's a concise list of the problems:
1. Half-band filtering: The upsampling (interpolation) and downsampling (decimation) filters are often specified such that the frequency response is symmetric around fsh/4;0.5 (on a linear amplitude and frequency scale. fsh=2fs). The effect of this symmetry is that every second coefficient is 0, except for the middle one, which is unity. This halves the number of computations necessary to implement a filter of the same steepness. However, the symmetry means that the response at 0.5fs is -6dB, a far cry from the total rejection called for by the nyquist theorem. Most if not all of PCM implementations actually do not conform to the nyquist criterium.
Now here's some funny psychology. Looking at the data sheets of converters you'll find that the passband goes to 0.4535fs, so the stopband starts at (1-passband) ie. 0.5465fs. At 44.1kHz, this translates into twentythousand point zero Hz. Apparently the belief must have been that as long as everything is well until 20.0k, all is cool and nobody will care about aliasing between 20.0k and 24.1k.
2. Equiripple filtering: These days chips are fast enough to run longer filters, so the half-band filter has gone somewhat out of fashion. One thing in which half-band filters are seen to be something of an overkill is in passband ripple. Due to the symmetry, a HB filter that has -100dB stop-band rejection has +/-0.0001 dB flatness in the passband. Not requiring a halfband characteristic but a looser passband spec (such as 0.01dB) will make the filter shorter in terms of time. All coefficients are now nonzero, but the total number of coefficients can be shrunk considerably, resulting in a reduction of group delay. Group delay is a serious problem in multitracking where a musician needs to hear herself play in the mix while tracking.
The result is that the filters are now commonly spec'ed as 0.01dB (or worse) flat until 20.000kHz (re 44.1kHz), stopband from 24.100k, because that was the accepted practice till now, no? Requiring a nonaliasing response would either mean shrinking back the passband to about 18kHz (a commercial no-no) or increasing the filter length (where the exercise was all about shortening it). The alias band thus remains status quo at 0.4535fs to 0.5465fs.
Why am I making such an issue of this flatness affair? Well, this periodic ripple corresponds to two small secondary spikes (echos) in the impulse response. One at the start, one at the end. For a filter that's flat to +/-0.01dB, these spikes are -66dB, and are a serious threat to stereo imaging and produce pretty obvious time smear (well, obvious to us perfectionists - aren't we a flea in the fur of the industry?)
3. Sharp filtering
A sharp filter has the effect of psychoacoustically enhancing the corner frequency, resulting in a kind of unnatural brightness or roughness. This effect is to my knowledge the most innocuous of effects associated with PCM. Also, making the filter steeper from what it is to begin with has no additional "brightening" effect.
How to attack the problems:
1. The aliasing problem can be solved at once, anywhere in the audio chain, using a single lowpass filter that enters stop-band before the alias band ie before 0.4535fs. A good place to do this is at reproduction or before final dithering. This means that halfband filters as in 1 may be used throughout without deleterious effects. I find that running a CD through an ultrasteep filter (pb to 18.5kHz, sb from 20kHz, eliminating all aliases that were created anywhere in production) results in an improvement in contrast, depth and precision of the stereo image.
2. The echos due to equiripple filters cannot be removed except by employing surgical precision on a case-by-case basis (like I'm doing to solve this problem on a TI SRC4192 - putting a DSP before and after the chip to compensate for the echos in the int and dec filters. In short, equiripple filters must be avoided at all cost. Simply specifying a higher sampling rate and a looser filter has the same effect of shortening group delay, but this implies that chips specified in this way are no longer suitable for the low sampling rates of 44.1k and 48k.
3. As we're already increasing sampling rate in view of the group delay problem, we can push it further to obtain a more natural roll-off after 20kHz.
Recipe for perfect PCM:
Specify a sample rate well above twice the audio band, e.g. 192kHz (lower is arguably acceptable too but since we're slowly standardising on 192 who cares)
Specify all interpolation/decimation filters as halfband, 0.4fs to 0.6fs transition.
Put exactly one non-halfband lowpass filter in the chain (e.g. at replay or before final dithering) that enters stopband at 0.4fs. Specifying its passband at 20kHz will allow for a very smooth roll-off and hence very short and practically ringing-free impulse response.
A more detailed paper concerning the aliasing and equiripple problems can be found at
www.nanophon.com (read them all - excellent stuff).
A description of a method to calculate the "final lowpass filter" is AES preprint 5822, by Peter Craven.
Gardo:
Every recording medium has fundamental flaws. CD tried to address some of the fundamental flaws of vinyl and introduced flaws of its own. I'm sure this will be true no matter how much net improvement is realized in recording and playback. The question for me is whether SACD (generally speaking) sounds better than CD (generally speaking). On my modest system, it certainly does.
BP:
On my less than modest system it does too. I'm only saying that of all the possible solutions that could be formulated for the deficiencies of CD, DSD is the least intelligent one.
Gardo:
BP says that "SACD is here to stay." Do we think that's a step forward or a step back?
BP:
Sonically it's a step forward. In terms of practical usefulness it's more like putting the world's population on a spaceship to colonise a new planet without first checking if there's water and oxygen on it, and if the conditions on earth were really so bad we needed to leave it.
Gardo:
If I'm following your recommendations correctly, it seems that the second sentence should end "do not produce audible artefacts." Is that right?
BP:
Indeed: "do not". I've made the same typo in a different discussion not so long ago and it made things a bit confusing.
Gardo:
Also, I'm gathering from your remarks that you do not believe there is any reason to have a passband greater than 20khz, so I take it you do not agree with those who say there's extra information up there that we need to be able to reproduce. Is that correct?
BP:
I was a bit unclear there...
The passband only needs to be flat up to 20kHz, but it shouldn't drop off after that - there's quite a lot of sonically relevant information above 20k. In the case of the 192kHz system I proposed, the stopband (-80dB or better) would start at about 77kHz. The -3dB point would lie at 36kHz, -6dB at 43kHz, -10 at 49kHz. This means a lot gets through and it's a nearly perfect compromise between impulse response and bandwidth.
Also keep in mind that this filter should be applied only once over the entire signal chain. If it's done in the customer's player, he can even chose between this and a flatter (but "ringier") response. My hunch is that the slow rolloff option is the most sonically transparent.
I have also included a couple of other questions that some of you might find of interest:
TH:
With regards to SACD playback, I notice that some multiformat players convert DSD to PCM before output. Does this conversion degrade the SACD sound in any audible way?
BP:
The conversion to PCM in these players is done by chips by NPC (2 versions available). They can downsample to 8, 4, 2 and even 1fs. How badly the sound is affected depends very much on which chip/setting is used. Someone here has ordered samples so we can measure the filters. The spec says nothing about them. A lot depends on how well the filters are implemented (see last mail). Probably the 8fs version is relatively innocuous. Funnily enough this is done in order to allow the DSD signal to be reproduced by Burr-Brown multibit DAC chips.
SACD players exist in an enormous variety, up to and including devices (usually the cheapest ones) that downsample to 44.1kHz. Also their DVD signals are reduced to 48kHz first. This is why indeed you have to be careful about buying a cheaper player.
The audiophile brands each (or at least many of them) have their own converter philosophy, which warrants attention to detail but not necessarily knowledgeability :-)
I personally use a first generation SACD player (Marantz SA-1) which happened to be lying around here. I modified it to deliver the DSD at three BNCs at the back and have my own DAC to convert it to analogue. The same DAC will take 192/24 (but it does not yet have all the fancy filters present - next version) so by the time I can knock off a DVDA player somewhere I can use that too.
TH:
You seem to be an audiophile as well as an engineer. What is you opinion of vinyl as a delivery format for analog master tapes? The re-issuing labels like the late DCC, Classic Records, Sundance, etc. seems to put out some really good sounding releases that, unlike the old days, do not seem to be compromised (compressed or limited) in any way?
In terms of sound quality I believe both SACD and DVDA to be good formats for releasing remasters on. As said before, my problem with SACD lies in the practical side of affairs.