To DSD or not to DSD?

[Miska]

5MHz? Why? I am not listening trough a short-wave receiver.

 

Because you tout so much about ultrasonic noise, so I wanted to see how your DAC manages. IOW, quality of your digital and analog filters.

 

When your device has -90dB THD+N then it doesn't matter that you measure it at -144dB

 

Pretty bad figure, just 15-bit worth.

 

Also, your (and mine) THD+N measurements don't represent the transient response to a bunch of sine waves or different amplitudes (aka music). I personally don't listen to sine waves. That's where the Delta-Sigma fails. It's error feedback lop has similar drawbacks on transients - similar with what happened in analog compressors.

Also decimation happens only in Delta-Sigma, people here tend to blame 'PCM' for that.

 

With PCM ADC you either have to use and keep high sampling rate, or use oversampling and then decimation filtering to RedBook. There is no way to make NOS PCM ADC for RedBook that would have good performance.

 

Timing performance is much worse problem for PCM. RedBook even theoretically cannot have rise time better than 23 µs because of the 44.1 kHz sampling rate and thus limitation to 22.05 kHz bandwidth. When you run the signal through oversampling filter things become more complex, for example the DF1700 rings for 3.5 ms, so essentially the trasient becomes blurred to 3.5 ms.

 

Since DSD doesn't need steep anti-alias filters on ADC side, nor decimation or interpolation filters at playback side, it has 1.4 MHz bandwidth (Nyquist frequency). Full-scale rise time of DSD64 is 8.5 µs which is quite a bit better than PCM. And already -6 dBFS signal rise time for DSD64 is just half of that 4.25 µs.

 

See below a PCM61 THD+N - a 16 bit signal (CD), 44.1kHz, undithered.. It just doesn't paint the whole picture.

 

My DACs have better performance even with just DSD64. You also seem to have PSU related jitter problem, seen as 120 Hz sidebands of the 1 kHz signal.

 

(you really should use dither)

[Audio_ELF]

(you really should use dither)

Hey I dither all the time... I can't make up my mind if either of you are posting anything of any consequence! :-)

 

Eloise

[mav52]

5MHz? Why? I am not listening trough a short-wave receiver.

I was just wondering what DAC are you using these days, thanks

[mav52]

Hey I dither all the time... I can't make up my mind if either of you are posting anything of any consequence! :-)

 

Eloise

 

DSD or not to DSD sounds like a dither

[Audio_ELF]

DSD or not to DSD sounds like a dither

A bit like Elton John too... Toupee or no toupee!

[John Siau]

Back on the topic 'To DSD or not to DSD' (from advertising posts), I 'll share this from DAD documentation:

 

 

Note that moving from DXD to PCM176.4 there will be no degradation. From DXD to DSD (any flavor) to me is raising the question 'why would you do that?'.

 

The above graph is misleading and incorrect. While some 176.4 kHz PCM converters may exhibit the rise in ultrasonic noise shown in this plot, most do not. For example, the idle channel spectrum of the Benchmark ADC1 and ADC16 is flat to the Nyquist frequency (see: ADC1 USB - Performance | Benchmark Media). The DXD converters used for the demo recordings played at the workshop at AES 2013 NYC also had a noise floor that was flat to Nyquist (flat to 176.4 kHz). The noise spectrum of the converters used for the DXD recordings was discussed in detail in the workshop. Like the Benchmark converters, these DXD converters had no rise in ultrasonic noise. In contrast, the noise spectrum of the DSD and DSD 128 is an unavoidable byproduct of the 1-bit quantization and noise shaping.

[Miska]

The above graph is misleading and incorrect. While some 176.4 kHz PCM converters may exhibit the rise in ultrasonic noise shown in this plot, most do not.

 

Yes, it is also worth noting that in many cases the noise floor is artificially flattened out by having the digital decimation filter roll-off early. So many ADC chips don't go flat to 88.2 kHz at 176.4k sampling rate.

 

P.S. I am seeing passband ripple in the ADC1 wideband frequency plot, due to digital filter?

[Miska]

for example the DF1700 rings for 3.5 ms, so essentially the trasient becomes blurred to 3.5 ms

 

Sorry, my math mistake due to posting in hurry, the ringing is of course half because first filter is computed at 2x rate and thus the ringing is very typical 1.73 ms. DF1700 is a traditional design and has three cascade filters.

[Hiro]

Yes, and exactly!

 

The most amusing part of all this is DSD hater Mark Waldrop (of AIX) just discovered DXD, and blogged about it. While he loves the numbers, he's mystified with the samples he downloaded from 2L that show a "purple haze" in the ultrasonic frequencies on his spectrograph.

 

He's clueless.

 

What About DXD? Surprise! | Real HD-Audio

 

Funny that he concluded that 44.1kHz CD might actually be sonically superior to DXD.

[John Siau]

Yes, it is also worth noting that in many cases the noise floor is artificially flattened out by having the digital decimation filter roll-off early. So many ADC chips don't go flat to 88.2 kHz at 176.4k sampling rate.

 

P.S. I am seeing passband ripple in the ADC1 wideband frequency plot, due to digital filter?

The ADC1 is a non-aliasing design - the pass-band ends just below Nyquist. Most converters have filters that are centered at Nyquist, and therefore some aliasing is allowed. The pass-band ripple that you are seeing is shown on a very expanded scale (0.05 dB/division). http://benchmarkmedia.com/adc/adc1-usb/performance

[Miska]

The ADC1 is a non-aliasing design - the pass-band ends just below Nyquist. Most converters have filters that are centered at Nyquist, and therefore some aliasing is allowed. The pass-band ripple that you are seeing is shown on a very expanded scale (0.05 dB/division). ADC1 USB - Performance | Benchmark Media

 

I meant the decimation filters inside ADC chip that convert from high speed sigma-delta to PCM.

 

I noticed the scale, but +-0.02 dB surprisingly large passband ripple, about the same as CS5381 has (0.035 dB p-p). For example the internal filter of PCM4222 has +-0.00015 dB ripple. And PCM4202 has +-0.005 dB.

 

I don't mind noise floor increasing 10 - 20 dB at 90 kHz, but I do mind passband ripple starting already at 2 kHz...

 

So far I have not seen any value keeping flat noise floor up to 100 kHz, instead I use noise shaping also with PCM, because that way I can gain tens of dB's of extra DNR in audio band while trading a bit of DNR in ultrasonic range. Plus, once you feed it through DAC it actually improves linearity especially with R2R ladders.

[Sam Lord]

Message deleted by writer.

[SoNic67]

Since DSD doesn't need steep anti-alias filters on ADC side, nor decimation or interpolation filters at playback side, it has 1.4 MHz bandwidth (Nyquist frequency). Full-scale rise time of DSD64 is 8.5 µs which is quite a bit better than PCM. And already -6 dBFS signal rise time for DSD64 is just half of that 4.25 µs.

True, but that's just for ONE bit. You need to work out many of those, feedback them, to generate the right signal. For me that's' not enough resolution.

[Miska]

True, but that's just for ONE bit. You need to work out many of those, feedback them, to generate the right signal. For me that's' not enough resolution.

 

No, one bit at DSD64 is 354 ns.

 

That feedback is really short compared to number of taps in digital decimation and oversampling filters used with PCM. So PCM filters lower the resolution much more than noise shaping filters. IOW, the filter order for PCM filters is in hundreds while orders for noise shaping loops are most of the time less than ten. I'm using ninth order noise shaper for PCM.

 

Most of the time, noise shaping order is much lower than oversampling factor, so it doesn't reduce the timing resolution at all. But it increases level resolution more than 10x.

[SoNic67]

Well, good, because you need that. 10x means how many bits added to the one? 3-4?

[gmgraves]

Have you ever listened any music ADCed and DACced by SONY PCM-1610/1630?

I have never. I wish I would could listen it!

 

Most early CDs were "mastered" using a Sony 1600-series ADC/encoder. That should tell you all you need to know about the

sound of these clunkers. Filled with 709/741 vintage op-amps and interstage coupling via aluminum electrolytic capacitors virtually

guarantees that the results were at least part of what gave digital its initial bad reputation in audiophile circles.

[Miska]

Well, good, because you need that. 10x means how many bits added to the one? 3-4?

 

No, increased sampling rate & noise shaping & same amount of bits. Here is one example, upper channel (1) is 16-bit PCM with noise shaping, lower channel (2) is 24-bit PCM with flat TPDF dither. Even though the first channel has 8 bits less, the audio band performance is still not that far from the 24-bit one.

 

We can keep going further and further on this by increasing sampling rate and reducing bits.

 

[Bunpei]

 

He's clueless.

 

What About DXD? Surprise! | Real HD-Audio

 

Is his web site currently being hacked by any reason?

[One and a half]

Is his web site currently being hacked by any reason?

 

Well and truly hacked, oh dear. I would be worried if I bought something...I did about 3 years ago.

[Sam Lord]

Have you ever listened any music ADCed and DACced by SONY PCM-1610/1630?

I have never. I wish I would could listen it!

 

It sounds like a bog-standard CD player:

 

Most early CDs were "mastered" using a Sony 1600-series ADC/encoder. That should tell you all you need to know about the

sound of these clunkers.

 

And to an engineer these machines were reviled for their sound. The late, great Denny Purcell of Georgetown Masters got the first 24bit/96kHz Sony hard-drive ADC+DAC recorder/playback device in existence, and his life improved dramatically after that.

 

First of all, don't bring NOS as argument. It's just not how PCM is designed to work... at 8x OS, you can get results from CD sources that sound better than DSD.

 

The first part makes sense, but 16bits cannot compare to well-managed 24 bits PCM (even at 44.1kHz) or DSD. A few seconds listening to playback of massed instruments or voices makes that plain.

 

Even PCM61 sounds better than any of your precious D-S DAC's. You think that linearity at 24 bit is important when actually is not. D-S design sacrifices musical quality for sake of good static pure sinusoidal numbers. That's just marketing.

 

That "marketing" has apparently fooled the best ears on the planet.

 

I personally don't listen to sine waves...

 

You've been doing that your whole life, we all have.

[Bunpei]

Since they talk about analog FIR filters for DSD and such, I suspect they use TI/BB DSDxxxx DAC chips like before in other products.

 

On his blog page, "kou" is showing some photos of his SONY HAP-Z1ES HDD player.

He identified TI/BB PCM1795 DAC chips on the DAC board.

 

マルチチャネルDAC製作記: SONY HAP-Z1ES のDACはカスタムチップ？　3

(This page is written in Japanese.)

[SoNic67]

You've been doing that your whole life, we all have.

That's the biggest fallacy. Music in not ONE sine wave. It might be approximated with an INFINITY of sine waves, but that's not what the tests are doing, is it?

[Miska]

That's the biggest fallacy. Music in not ONE sine wave. It might be approximated with an INFINITY of sine waves, but that's not what the tests are doing, is it?

 

Oh, we can test with impulses that contain all frequencies at once.

[SoNic67]

Well, you can try. And what do you see? I don't listen to pulses either. Pulses are just a surface indicator, good for a start. Same with THD+N. Same with SNR. They are not the ultimate result and pushing them over a certain limit doesn't translate in bettter sound quality. We don't have bat ears, but we have a brain that can detect subtle changes in harmonic content and it's variations. Noise shaping affects that in certain moments. ONE static sine wave is not one of those moments tough.

 

The issue is very old - since the time of the analog compressors-expanders. Re-iterated again with mp3 conversion. Present in DSD conversion with feed-back circuits that are inter-dependent. Timing that the real-time processing of audio is adding to the decision-making blocks.

 

This affects excursions from high amplitude signal to low amplitude ones. Mix of high amplitude/low frequency components with low amplitude/high frequency ones will create a different noise shaping then when only the low level/high frequency are present. Harmonics spectrum will vary during this excursion (different dither spaces) and brain is more likely to pick up on that then a constant harmonics spectrum (like PCM non-dithered).

[wisnon]

Sonic67,

 

The day you hear DSD128 done right is the day you will change your mind. I will welcome you into the fold with open arms.

 

Come on in, the water is fine!

Hahahahaha