A couple of riffs on the notion of "bit perfect"...

[spdif-usb]

That paper is just outdated and incomplete/incorrect. SDM signal processing has proceeded quite a lot since.

Now I am very curious to find out how the non-linearities of the DSD-encoded signal can be fully dithered out.

And I know how I'm doing it in practice... That's the way Sonoma does it, and my way is different.

Right. But then, how do we convince HDTracks to fire the guy and hire you instead? :grin:

That's the way probably all DSD DACs work, from practical point of view you can consider PDM being equal to PWM. The difference is only theoretical. I would say probably all SDM and PCM DACs are sample-and-hold type. For ADCs it may actually vary up to being 50/50 or something.

I kind of already figured that PDM was practically the same as PWM. Both are 1-bit after all. Anyway, thanks for clearing that up.

When the data is 30% less to begin with, the remaining compression ratio brings the total bandwidth savings to roughly equal of FLAC. Not that it would matter in practice, audio is so low bandwidth anyway that dealing with it as uncompressed is not a problem.

Perhaps it can be roughly equal of FLAC as you suggest, but doesn't 24/192 offer more resolution, or headroom, to work with in terms of processing?

(good quality full-HD video with H.264/AVC compression is 24-30 Mbps at least peaking around 50 Mbps with pans)

This brings up yet another interesting question IMO. If I am wrong about DSD so that it actually does offer some advantages over rate-converted PCM, why doesn't the Blu Ray format make better use of them? With the number of channels per audio stream as well as the number of selectable audio streams that can be provided on a single commercial Blu Ray disc seeming to never suffice, surely one would expect the company that should be considered technically very close to the top of both formats (Sony) to have pushed DSD as the primary standard for Blu Ray?

[Jud]

Now I am very curious to find out how the non-linearities of the DSD-encoded signal can be fully dithered out.

 

 

Well, how low is "fully"? The 2001 paper in essence says since the noise in DSD can't even in principle be dithered away, while the noise in PCM in principle can, high end A/D and D/A manufacturers and consumers should henceforward concentrate on PCM and abandon DSD. That of course didn't happen.

 

Now although the noise in PCM may in principle be able to be fully dithered out, are you aware of any application where a perfect lack of noise is claimed? Certainly the electrical/thermal noise of the equipment would constitute a lower bound, so in practicality a state of no noise whatever is impossible. The corollary of this is if DSD noise shaping pushes audible noise down into that same almost vanishingly low (but not perfectly absent) area of which PCM is capable, then for practical purposes DSD is an equivalent technology with respect to noise.

 

Two links with reasonably current information on performance of DSD vis-a-vis noise are dsd (see figure 2) and http://www.esstech.com/PDF/sabrewp.pdf (see noise figures mentioned throughout the paper and performance graphs at the end). It seems to me the noise performance shown in these references is fairly comparable to that obtainable with high-res PCM.

[spdif-usb]

It appears to me you are confused about dither. Noise cannot be dithered away because, as a matter of true fact, dither is all about adding noise. Quantization errors are effectively masked by the noise that gets added by the dithering algorithm, and as a result from this it becomes possible for the temporal resolution of PCM to be made infinite. The resulting noise floor can become audible if the bit-depth of the signal is only 16-bit, but the audible quantization errors of the 16-bit signal are alot more detrimental to the sound than the noise is.

The graph in that paper on DSD shows an FFT plot of the noise floor for a DSD signal, and the individual strands of noise can be seen. However, if you created the exact same type of FFT plot of the noise floor for a PCM signal, you would see that there is a rather huge difference between marketing and marketeering. :cough:

 

Well, how low is "fully"? The 2001 paper in essence says since the noise in DSD can't even in principle be dithered away, while the noise in PCM in principle can, high end A/D and D/A manufacturers and consumers should henceforward concentrate on PCM and abandon DSD. That of course didn't happen.

 

Now although the noise in PCM may in principle be able to be fully dithered out, are you aware of any application where a perfect lack of noise is claimed? Certainly the electrical/thermal noise of the equipment would constitute a lower bound, so in practicality a state of no noise whatever is impossible. The corollary of this is if DSD noise shaping pushes audible noise down into that same almost vanishingly low (but not perfectly absent) area of which PCM is capable, then for practical purposes DSD is an equivalent technology with respect to noise.

 

Two links with reasonably current information on performance of DSD vis-a-vis noise are dsd (see figure 2) and http://www.esstech.com/PDF/sabrewp.pdf (see noise figures mentioned throughout the paper and performance graphs at the end). It seems to me the noise performance shown in these references is fairly comparable to that obtainable with high-res PCM.

[Miska]

The graph in that paper on DSD shows an FFT plot of the noise floor for a DSD signal, and the individual strands of noise can be seen. However, if you created the exact same type of FFT plot of the noise floor for a PCM signal, you would see that there is a rather huge difference between marketing and marketeering. :cough:

 

What you really need to look at is the analog-filtered DAC output up to fs frequency, and it's linearity and noise levels at wanted audio frequency band. Pay special attention to any frequency components that correlate with the original signal.

 

For example compare noise levels and low level linearity of multi-bit PCM ladder DAC and DSD DAC in 0 - 20 kHz frequency band and you'll see that DSD DAC can actually achieve better low level linearity and lower noise. This is because there cannot be any bit-to-bit non-linearity that plagues PCM converters.

 

DSD has the non-correlated ultrasonic noise on purpose to linearize the audio band, while PCM ladder DAC has horrid correlated de-linearizing alias frequencies in the ultrasonic range. As result if you get anything back from DSD noise as intermodulation products back to audio band as intermodulation products, those are just random plain noise increasing the noise floor a little bit across the board. While with PCM all the intermodulation products folding back to audio band are discrete frequencies correlated with the signal - becomes real audible distortion.

[Miska]

Now I am very curious to find out how the non-linearities of the DSD-encoded signal can be fully dithered out.

 

You can start with newer AES papers posted in response to that old one.

 

Right. But then, how do we convince HDTracks to fire the guy and hire you instead? :grin:

 

I'm a software/hardware engineer designing new tools, not a mastering engineer using the tools. So I wouldn't be very useful for HDtracks unless they need new tools. They didn't develop the Sonoma DAW either.

 

I kind of already figured that PDM was practically the same as PWM. Both are 1-bit after all. Anyway, thanks for clearing that up.

 

Sometimes I find it hilarious when some people advocate class-D amps while being vocal about DSD noise, etc...

 

Perhaps it can be roughly equal of FLAC as you suggest, but doesn't 24/192 offer more resolution, or headroom, to work with in terms of processing?

 

No, not necessarily. Practically always 24/192 is sourced from SDM bit stream and has been spoiled in time domain by a brickwall filter. Most SDM-to-PCM conversions in ADC chips leave a lot to wish for, so it would be better to leave it to some more powerful DSP if converted at all.

 

This brings up yet another interesting question IMO. If I am wrong about DSD so that it actually does offer some advantages over rate-converted PCM, why doesn't the Blu Ray format make better use of them? With the number of channels per audio stream as well as the number of selectable audio streams that can be provided on a single commercial Blu Ray disc seeming to never suffice, surely one would expect the company that should be considered technically very close to the top of both formats (Sony) to have pushed DSD as the primary standard for Blu Ray?

 

Studios have their tools already for PCM and in any case movie audio is practically controlled by Dolby and DTS. It is not about what would be technically best thing. There are lot of examples when technically best solution didn't win the market, for what ever reason.

 

In any case, HDMI supports DSD up to at least 192x64 rate and 8 channels...

[Jud]

It appears to me you are confused about dither. Noise cannot be dithered away because, as a matter of true fact, dither is all about adding noise. Quantization errors are effectively masked by the noise that gets added by the dithering algorithm, and as a result from this it becomes possible for the temporal resolution of PCM to be made infinite. The resulting noise floor can become audible if the bit-depth of the signal is only 16-bit, but the audible quantization errors of the 16-bit signal are alot more detrimental to the sound than the noise is.

The graph in that paper on DSD shows an FFT plot of the noise floor for a DSD signal, and the individual strands of noise can be seen.

 

Dither is a paradoxical thing. Yes, it adds noise. But it does so to lessen the level of the most noticeable noise. As you say, it masks quantization errors. Or as Wikipedia says: Dither should be added to any low-amplitude or highly-periodic signal before any quantization or re-quantization process, in order to de-correlate the quantization noise from the input signal and to prevent non-linear behavior (distortion); the lesser the bit depth, the greater the dither must be. The result of the process still yields distortion, but the distortion is of a random nature so the resulting noise is, effectively, de-correlated from the intended signal.

 

So no, I'm not confused. When I mentioned "dithering away" noise, I meant the more noticeable noise that dither is added in order to get rid of ("de-correlate the quantization noise from the input signal and to prevent non-linear behavior (distortion)"). This can't be done in the same way with DSD. But for DSD, noise can be "shaped," essentially moved around in frequency, which is what Miska mentions as "non-correlated ultrasonic noise on purpose to linearize the audio band."

 

However, if you created the exact same type of FFT plot of the noise floor for a PCM signal, you would see that there is a rather huge difference between marketing and marketeering. :cough:

 

Is the difference between marketing and marketeering anything like the difference between mousing and Mouseketeering?