View Single Post
Old 20th October 2006
Lives for gear
T.RayBullard's Avatar

Thread Starter
Originally Posted by lagoausente View Post
What´s advantage on record on DSD? As I understood, you need to convert it to PCM to can edit it..
some desks(like the sonoma workstation) edit DSD as is..and Sadie, I believe..and Pyramix(or does that one downsample)

as for DSD..the benefit is the SOUND...just loads better..IMHO..of course.
DSD Vs PCM from the Chief Engineer at Phillips

Studio people who have compared the live misc feed to DSD and PCM say that DSD is much better and they cannot tell the two apart. If, as you say, it is flawed, why are the studio engineers all for it. As someone on the forum said, it is analog without all the problems.

1) Sound of DSD:

The HF noise and the low-level nonlinearities of DSD do not get in the way of the sound as we hear it. However, it is impossible to build a production chain using the format.

DSD is in its place in one (1) application: Mastering from analogue. When the recording chain is completely analogue, you can feed the audio from the analogue mastering into a DSD A/D converter and cut that signal straight onto a disc without any further processing. It is in this application that DSD can be viewed as pretty transparent. When you convert the signal twice, however (such as when using a DSD recorder as the tracking medium), the second conversion is no longer transparent, due to the HF noise present in the source signal hitting a second analogue deltasigma modulator.

In this I have a serious gripe with AudioQuest. Before DSD, they tracked onto a 15ips 1/2" two track. These tapes were first used to master to ordinary CD, later again to SACD for a reissue, which indeed was a sonic forward step. However, they then switched to using a DSD recorder for tracking as well. Grundman then mastered it as usual (ie through an analogue mastering studio) onto the DSD master recorder.
The recording guys may well have found the DSD recorder better than their analogue two-track in a dry shoot-out, so you can't blame them for having made that choice. The resulting SACD releases however, are below anything. What's worse, there's no way of salvaging the recordings for a better-sounding reissue (unless they had the analogue two-track running as a backup). As an example, get hold of the classic "BluesQuest" sacd, which was made from analogue tapes. Then compare Doug McLeod's "Whose truth whose lies", where a DSD machine was used. If you're a bit of a sensitive person you'll run away screaming.

I think we could say that DSD is analog with a few extra problems. Serious ones.

2) Total transparency:

Apart from that I consider any claim of "not being able to discern between a live feed and DSD" as something of a hyperbole. While flicking a switch during the music will indeed not reveal any serious deficiencies, more controlled listening (e.g. listen - rewind - switch - listen) will certainly get you hearing a lot more. I can't even insert a unity gain stage (low-noise low-distortion etc) in the signal path without hearing degradation, let alone a pair of converters.

3) Jitter:

When a signal is fed straight from an ADC into a DAC, they share the clock. When you sample a signal with some jitter, and you reproduce it with the *same* jitter, the jitter has NO impact on the sound (jitter which impacts the sound most, incidentally is of the LF kind, which does not affect noise specs, and which is typically not attenuated by PLLs). With DSD this works - the delay between ADC and DAC is nearly zero. With PCM it doesn't, because there is a delay of several milliseconds between the two, meaning sampling and reconstruction see different jitter. This means that a live vs DSD will always sound more transparent than when you take the digital signal to tape and replay it.

4) PCM implementation issues:

On you can read a number of the late Julian Dunn's excellent papers on how compromised implementation of digital filters account for many of the deficiencies noted with PCM. These can be readily solved with due care for details and at the expense of only mildly increased computational burden.

How do you feel about the DSD workstations? The existing ones, I believe, are all using DSD-wide, or PCM-narrow.

The workstation from Merging is in itself OK in that the data is kept in 352.8kHz/32bit floating until it is written to an sacd master file. Currently the incoming data is DSD, causing a "two DSD conversions" problem (although apparently in the digital domain the sonic degradation is less than when it happens in the analogue domain). If you can somehow get the data in through a non-DSD converter (preferably at 352.8kHz/24bit or so), it would become a digital equivalent of the "ideal analogue chain" in that only one deltasigma coder is in the signal path (when writing the master). That would be fine.

If you do have to edit a recording which is already in DSD, it can be done in a sonically transparent way by using the "transport" mode of the workstation. In this mode, the DSD data is preserved exactly except during the edits (crossfades). The two conversions problem would be restricted to those edits only. Of course, any form of processing, such as level change, EQ or mixing, is not possible in this mode. It's really just cutting and splicing.

The workstations from the Sony camp (I don't know names) use 2.8224MHz, 8 bits, still noise shaped. This still has the same problem as reconverting to DSD every time, but reduced by 48dB. Still I wouldn't recommend it. The Merging approach is the most suitable one.

You mention the use of a non-DSD converter (preferably at 352.8kHz/24bit or so). EE Dan Lavry argues that the bigger the numbers the less accurate the conversion is. He says that there really is no need to go beyond 24/96 (48k bandwidth) and anything more will not result in a more accurate signal but rather more noise and distortion.

Dan is correct in that as the sampling frequency is increased, the available signal to noise ratio inside the nyquist band decreases. However, when you keep the bandwidth across which you are measuring SNR constant (e.g. you measure noise across 20kHz) and when noise shaping is used, this trend is often reversed. DSD is precisely a case in point. I have one converter here (homegrown discrete circuit) that puts out 1 bit at 2.8224MHz. Measured across its nyquist bandwidth (1.4112MHz), its SNR is useless, well below 6dB. However, taken across 20kHz, it delivers the full 120dB. When you push the sampling rate too high, performance will again resume a downward trend.

Dan's converters are multibit, non-noise shaping. Such converters will not even hold their precision at constant bandwidth when sampling rate is increased. Since his converters have ultra-low noise as their hallmark, he's quite right to maintain a reasonable sampling rate.

When I propose to use a 352.8kHz converter, in practice that would be a noise shaped converter designed to offer maximum SNR performance over as wide as possible a band, but not up to 176.4kHz (not feasible). It would still have a "noise shaping tail" in the nyquist band, but much less so than DSD. At the final DSD conversion stage, the noise of the ADC would be negligible compared to the DSD noise, so the output spectrum would be pretty much the same as that of an analog signal converted to DSD in one go.

A design which is in the work at home uses 64fs, 16level PWM with 7th order noise shaping. This would offer up to 129dB SNR (limited by amplifier noise) over 80kHz, which is 4 times as wide as DSD. This would insure all the flexibility of PCM while a later conversion to DSD is not compromised.

How do you feel about DSD as an archiving format? Is the transparency good enough for Sony's precious analog master tapes? And what makes DSD ideal for archiving as opposed to PCM? Or does it matter? I read Neil Young chose hi-res PCM format for his masters.

When sony came up with DSD as an archiving system there was hardly even 96kHz PCM around. If at that time they had some old tapes to archive before they fell apart, DSD was the best available. However, since DSD is a liability in terms of processability, archiving to DSD now is no longer a good idea and use of 192/24 is warranted instead. Since SACD is probably here to stay we should view DSD as strictly a release format, in the same way as we didn't produce on vinyl, but music was brought to the home on it.

What do you say when you hear audiophiles make comments that DSD has all the "air" and "smoothness" of analog?

In my own experience, high speed PCM also produces the air DSD has, while the "digital glare" of some PCM can even be solved at low speeds. It is caused by the narrow alias-band that is present between 20 and 24.1kHz. Removing this band prior to playback restores naturalness and focus.

I don't suspect DSD of any "euphonic" effects, although, who knows, the HF noise?

Admittedly, the DSD camp has been able to mobilise more audiophile designers (folks like Ed Meitner), resulting in the analog circuitry of the best DSD converters sounding better than that of most available PCM converters. Doing a straight shoot-out is actually quite challenging technically, as the two formats normally use different converters, necessarily producing a different sound. This is another reason for me to do this 352.8kHz converter, because its output can be converted to either DSD or 192/24 while compromising the performance of neither. This would finally allow a direct comparison.

Are the production chain problems insoluble, or not yet solved?

The "production chain problem" ie. the fact that signal quality quickly and irreversibly deteriorates as it passes through subsequent processing stages, is in itself not solvable. The root cause lies in the high HF noise level. This noise is indistinguishable from the wanted signal, so it cannot be stopped from accumulating every time a signal is converted to 1-bit. This is not to say there are no workarounds. However, the mere fact that such workarounds are necessary shows that the DSD format itself was misconceived.
1. "Production DSD". To use a 1-bit signal at 128fs or 256fs instead of 64fs. Especially in the case of 256fs, signal and noise no longer overlap. The signal bandwidth is specified, as before, at 70...80kHz. The quantisation noise only comes out of the analogue noise floor at 80kHz as opposed to 20kHz in the case of 64fs. Now, signal and noise can be separated, quite simply by filtering. It follows that every processing step still involves filtering to remove the HF noise from previous conversions and deltasigma modulation to recode the processed signal into 1-bit. This constitutes a considerable processing overhead.
2. PCM-narrow. To use 352.8kHz, 32bit floating point as the intermediate format. This is PCM by all means, but according to the listening tests carried out by Philips and Merging, the decimation and upsampling filters needed to convert to and from 2.8224MHz do produce audible artefacts. The advantage here is that this format is practical enough to maintain throughout production, store on hard disk etc. Converting from DSD to PCM-narrow is done only once, namely as the data comes into thr production chain. Conversion to DSD is only done once, namely at mastering. Moreover, the 352.9kHz data can be derived straight from a better-than-DSD AD converter (which means most of all present day converters), burdening the production chain with no or greatly reduced HF noise.
3. DSD-wide. 2.8224MHz, 8 bit. This format still requires all processed data to be noise-shaped back, but to 8 bits instead of 1. This can be done many times before noticeable headroom reduction or other adverse effects set in.