Dithering is a Mathematical Process - NOT a psychoacoustic process.

[j_j]

Forgive me if I've missed something, but isn't DeltaWave used for measuring signals rather than filter design? (i.e. For high quality audio processing applications, where the output signal will be ultimately "received" by the human auditory system.)

A filter is just a special kind of a signal that gets convolved by the actual "signal".

 

[j_j]

Perhaps you can clarify why certain test discs (CD) of mine use different (to one another) digitally generated (not from analog sign gen) prime frequencies for the high precision THD testing tracks?

It's hard to say from first principles. Something that hits all levels within a small number of cycles would always be good, of course. There are many reasons for THD signals to be multitone, including those that are high frequencies spaced at a particular (much lower) frequency apart. I think I just posted one of those here, yes? Such signals make IM distressingly obvious.

 

[j_j]

The question is whether the human auditory system can perform that statistical analysis for the detection.

Not unless you're talking about psychoacoustics, which you don't want to do. Make up your mind, m'kay?

 

[j_j]

GUIDANCE NOTE FOR RETAIL FUEL DISPENSERS (PETROL PUMPS) AND ROAD TANKER MOUNTED METER MEASURING SYSTEMS FITTED WITH STANDARD TEMPERATURE ACCOUNTING (STA) DISPLAYS

Not a link to be clicked on unless really curious. Rather, it's evidence that, when you fill the tank(s) of your vehicle(s) with petrol/gasoline, the amount dispensed is not exact, instead being within certain tolerances.

So, it just goes to show that dithering occurs, anyway. ;-)

e= h * nu

Dithering is inevitable at the smallest scales.

 

[Tks]

I don't think that came out quite right.

As for ESS in general, they are all about creating mystique around their chips without ever providing any solid information. They won't even give you a datasheet without an NDA, and even then it is, kindly put, incomplete. Why would anyone trust a single word they say?

A look at the history of ESS also doesn't inspire confidence:

• 1997: Mostly chips for PC (ISA) sound cards (comparable to SoundBlaster at the time), a couple of MPEG video decoders.
• 2000: Expanded PC audio line-up, addition of modems.
• 2002: Emphasis on modems and networking, addition of DVD processors and MP3 decoders. PC audio scaled back.
• 2004: Modems/networking and DVD processors. All pure audio products gone.
• 2006: Modems gone. Focus on DVD and low-end portable audio. Addition of image sensors for phone cameras.
• 2008: Image sensors gone.
• 2009: SABRE DAC and ADC line introduced, along with a VoIP processor.
• 2010: SABRE line expanded. TV tuner added.
• 2011: VoIP gone. Soundbar processor added.
• 2014: TV tuner gone.
• 2015: Focus on SABRE DAC. ADC and DVD products gone. Addition of headphone amplifiers.
• 2016: Voltage regulators added.
• 2019: USB audio dongle chips added.

This strikes me as a company focused on riding short-lived trends. Apparently, with the SABRE DACs they lucked into the audiophile market and decided to double down on that, abandoning everything else (as so many times before). Recently, they have latched onto the MQA fad (if you can call it that). We'll see how long that lasts.

If you want to learn from silicon industry insiders, look to Analog Devices and (to a lesser extent) Texas Instruments. Their datasheets and application notes contain a wealth of information without the hyperbole and mumbo-jumbo characteristic of ESS.

Even I as a complete noob to audio and laymen to virtually 90% of the discussion in this thread picked up on ESS's nonsense. Just the idiotic hump being such a pervading issue in most DAC's tested, stands to support this view of mine. EVEN IF the fault is with every single DAC maker being ignorant of a simple fix - the fact information of that sort isn't made apparent to all these designers, is what leaves a sour taste in my mouth with respect to ESS's approach.

I don't EVER want to see something like a chip designer be party to mystique building when it comes to offering products to designers or engineers. I can't imagine what stuipidity the world would look like if this translated to something like ARM or Intel or AMD talking about how their CPU offerings had a "magical" aspect that is beyond being shown on paper. Completely infantile nonsense.

 

[xr100]

A filter is just a special kind of a signal that gets convolved by the actual "signal".

Yes. But said (filter) signal is not of much use in the wider world if it can't be expediently stored for use outside of the analysis software referred to. ;-)

 

[xr100]

a "magical" aspect that is beyond being shown on paper. Completely infantile nonsense.

They have some patents which disclose some of their methods.

As I mentioned, the SABRE DAC white paper that I linked to in a previous post sets out how to conduct the "Noise vs. DC offset" test that they have used in generating plots comparing their products to those of competitors.

Alas, as I think some have noted elsewhere, the relative scaling on these charts is dubious. And, of particular relevance to this thread, in very small text the chart reads: "NO DITHER."

I don't think it's too hard to relate this to the "DF Metric" (that ultimately prompted the creation of this thread) and other aspects discussed in this thread...

Basically, a discussion on the "oddities" of Delta-Sigma conversion would be good...

 

[exaudio]

Perhaps you can clarify why certain test discs (CD) of mine use different (to one another) digitally generated (not from analog sign gen) prime frequencies for the high precision THD testing tracks?
...
And the Denon disc gives lower THD than the CBS CD-1 test disc (which was the test standard for years at all the audio magazines) on many of its high precision tracks too. Not only that, the Denon is balls accurate (benchtop freq counter) on all the spot frequencies, whereas the CD-1 isn't.

Huh. I would've thought if you've seen one 1kHz test tone you've seen them all--but I would've been wrong. I went and grabbed the two test CDs I have: A Denon Hi-Fi Check CD from 1992 and an Alpine Reference CD from 1988. I ripped both discs' 1kHz tracks to my hard drive for comparison. The Denon disc does have a 997Hz tone, but I didn't rip that one. Maybe I'll compare that one tomorrow.

The Alpine disc's test tones are licensed from Brüel & Kjær but the Denon's spectrum was far cleaner. Looking at the spectrums my guess is the Alpine disc must be a recording of an analog signal generator while the Denon must be digitally generated. And, as @restorer-john said, the "Denon is balls accurate." The track listing said 1kHz at -15dB and the ripped wav showed exactly that. The Alpine's 1kHz tone was listed at "0dB level" but the ripped wav was closer to -20dB.

Here's the spectrum plots of the two tracks compared using a couple different windows and a couple different apps:

Here they are in DeltaWave. Blue is Alpine, White is Denon. (Kaiser Window?)

Dirichlet Window:

Tukey Window?

Here's the spectrograms compared in Audacity. (Alpine on top, Denon Below)

If you rip the discs and post the tracks, we can perhaps figure something out.

I've attached the two wav files I ripped in case you're interested. I ripped them using Exact Audio Copy and both tracks checked out with the AccurateRip database. It's interesting to me that the Denon file zipped down much smaller than the Alpine. From an information theory standpoint I'm assuming that means there's less entropy there which for a test tone probably means less garbage.

 

[j_j]

Yes. But said (filter) signal is not of much use in the wider world if it can't be expediently stored for use outside of the analysis software referred to. ;-)

Now, yes, indeed, that's true, unless you only need the filter inside the software. But then that wouldn't be "the wider world".

 

[scott wurcer]

I can't imagine what stuipidity the world would look like if this translated to something like ARM or Intel or AMD talking about how their CPU offerings had a "magical" aspect that is beyond being shown on paper. Completely infantile nonsense.

That might not be literally true but many processors have undocumented op-codes. Intel has a list of benchmarks for common math operations using their compiler and processors vs gcc.

 

[xr100]

I've attached the two wav files I ripped in case you're interested. I ripped them using Exact Audio Copy and both tracks checked out with the AccurateRip database. It's interesting to me that the Denon file zipped down much smaller than the Alpine. From an information theory standpoint I'm assuming that means there's less entropy there which for a test tone probably means less garbage.

Thanks for uploading the files!

Here's some spectrograms from iZotope RX...

The Alpine (B&K) track with 1kHz component removed using the spectral editor in RX, then normalised:

The Denon test track with 1kHz component removed using the spectral editor in RX, then normalised:

The Alpine (B&K) test track further processed using RX's "Deconstruct" process, with "tonal" components removed:

The Denon test track further processed using RX's "Deconstruct" process, with "tonal" components removed:

So, yes, vastly more "garbage"/"stochastic" behaviour in the Alpine (B&K) test track.

That said, the Alpine disc is presumably fine for its intended application(s) (i.e. not for measuring D/A converters, etc..)

 

[pkane]

Here they are in DeltaWave. Blue is Alpine, White is Denon. (Kaiser Window?)

Click on the little gear icon near the top right corner of DeltaWave to bring up the configuration window. Check what Window and window size are selected under Spectrum settings:

 

[scott wurcer]

Thanks for uploading the files!

Here's some spectrograms from iZotope RX...

The Alpine (B&K) track with 1kHz component removed using the spectral editor in RX, then normalised:

View attachment 49126

The Denon test track with 1kHz component removed using the spectral editor in RX, then normalised:

View attachment 49128


The Alpine (B&K) test track further processed using RX's "Deconstruct" process, with "tonal" components removed:

View attachment 49129

The Denon test track further processed using RX's "Deconstruct" process, with "tonal" components removed:

View attachment 49130

So, yes, vastly more "garbage"/"stochastic" behaviour in the Alpine (B&K) test track.

That said, the Alpine disc is presumably fine for its intended application(s) (i.e. not for measuring D/A converters, etc..)

The Denon looks typical of an undithered test tone. The cyclical nature of the rounding puts the errors in tones rather than noise. The Alpine is a mess.

 

[xr100]

Alas, as I think some have noted elsewhere, the relative scaling on these charts is dubious. And, of particular relevance to this thread, in very small text the chart reads: "NO DITHER."

Noise shaping can do the same thing, but can create idle tones, and other interesting things. Noise shaping should be dithered, and they aren't quite the same thing. Strictly speaking, noise shaping does not linearize completely, but it does let you get below the lsb.

 

[mansr]

The Denon track has some leading and trailing silence. Inside this, the tone fades in/out over about 800 samples. This is making some of the spectrum plots above uglier than they should be. The remainder consists of a sequence repeating every 441 samples. The spectrum of this using an FFT size of 44100 and rectangular window looks like this:

A quick attempt at generating this signal from scratch using rounding doesn't quite match on all samples. This is probably simply due to different accuracies used in the calculations, and I'd rather not spend any time trying to find the exact recipe.

The Alpine track is rather nasty. It fades in/out over about a second, so that has to be trimmed. This is where it gets weird. Counting zero crossings puts the frequency at 1024.8 Hz. Most periods have 43 samples, a few 44. Poking at the signal, I noticed it had a growing DC offset (so not truly DC, for the pedantic). For a better look at this, I calculated a moving average with a window of 1334 samples (very close to 31 periods). Ideally, an average over an integer number of periods is zero. Instead, we have this:

WTF? How does this happen? Leaving that mystery unsolved, we shall instead take a look at the spectrum. This being one of those times when the FFT can't be exactly matched to the signal, we use a Dolph-Chebyshev window. The result:

WTF, again. Horrible harmonics, strange spurious tones, and a smattering of noise at the low end of the spectrum. The THD is -59 dB, THD+N -57 dB. This is much worse than just about any CD player. What exactly is this track supposed to test?

For those who prefer frequency on a log scale:

Is that 60 Hz power line hum?

This is clearly a recording of an analogue signal. Didn't they have computers back in '88? Were tone generators and ADCs really this poor? Why would anyone pay money for this?

 

[scott wurcer]

The Denon track has some leading and trailing silence. Inside this, the tone fades in/out over about 800 samples. This is making some of the spectrum plots above uglier than they should be. The remainder consists of a sequence repeating every 441 samples. The spectrum of this using an FFT size of 44100 and rectangular window looks like this:
View attachment 49149

A quick attempt at generating this signal from scratch using rounding doesn't quite match on all samples. This is probably simply due to different accuracies used in the calculations, and I'd rather not spend any time trying to find the exact recipe.

The Alpine track is rather nasty. It fades in/out over about a second, so that has to be trimmed. This is where it gets weird. Counting zero crossings puts the frequency at 1024.8 Hz. Most periods have 43 samples, a few 44. Poking at the signal, I noticed it had a growing DC offset (so not truly DC, for the pedantic). For a better look at this, I calculated a moving average with a window of 1334 samples (very close to 31 periods). Ideally, an average over an integer number of periods is zero. Instead, we have this:
View attachment 49147

WTF? How does this happen? Leaving that mystery unsolved, we shall instead take a look at the spectrum. This being one of those times when the FFT can't be exactly matched to the signal, we use a Dolph-Chebyshev window. The result:
View attachment 49154

WTF, again. Horrible harmonics, strange spurious tones, and a smattering of noise at the low end of the spectrum. The THD is -59 dB, THD+N -57 dB. This is much worse than just about any CD player. What exactly is this track supposed to test?

For those who prefer frequency on a log scale:
View attachment 49155

Is that 60 Hz power line hum?

This is clearly a recording of an analogue signal. Didn't they have computers back in '88? Were tone generators and ADCs really this poor? Why would anyone pay money for this?

I tried in Audition to make a 1K tone at 32bit FB and then reducing it without dither to 16 bits and the results were almost identical, the 100Hz spaced tones are as you say from the common divisor between 44100 and 1k. Possibly one starts with 64 bit floats and the other 32 bit? A few years ago I tried Matlab, Mathematica, Python, and our in house SPICE all probably compiled with the IEEE math library and the results were all exactly the same to 15 digits.

The other one looks like one of those hobby oscillators with a diode shaper for a sine wave and a horrible power supply.

 

[xr100]

A quick attempt at generating this signal from scratch using rounding doesn't quite match on all samples. This is probably simply due to different accuracies used in the calculations, and I'd rather not spend any time trying to find the exact recipe.

Checking a few periods, I counted 44 samples between zero-crossings.

44100/44 = 1002.272727... Hz.

Then in the tone generation process, whatever precision is used for π, and the trig function that's generating the waveform.

Anyway, here are some more spectrogram plots generated by iZotope RX.

1002.272727... Hz (64-bit float) @ -15dBr, truncated to 16-bits (File is attached):

Compared to the Denon test track:

WTF, again. Horrible harmonics, strange spurious tones, and a smattering of noise at the low end of the spectrum. The THD is -59 dB, THD+N -57 dB. This is much worse than just about any CD player. What exactly is this track supposed to test?

Totally crummy but leaving out the "licensed by B&K" moniker it's an ALPINE test disc. So, presumably it was intended for use in "Sound Offs" when measuring which system was capable of 155dB continuous at 1kHz, or 160dB. ;-)

For the kind of "informal" uses it's likely to have been used for, it's fine. Well, apart from a track ostensibly having a waveform amplitude at 0dBr actually being at -20dBr. It presumably could still be used with a voltmeter to adjust power amp gain or whatever...

Is that 60 Hz power line hum?

The 2nd harmonic is stronger:

(Note: Adjusted amplitude scale c.f. above plots.)

Harmonics can be seen at 180, 240, 300, 360 and 480Hz also:

This is clearly a recording of an analogue signal. Didn't they have computers back in '88? Were tone generators and ADCs really this poor? Why would anyone pay money for this?

In terms of the A/D conversion, I wouldn't be surprised if it's representative of the some digital devices of the era that were used in professional recording studios. For instance, the Publison Infernal Machine (1985), which was one of the first samplers that had enough RAM to store enough seconds of audio to make it usable for doing tasks such as quickly "flying in" (copying) vocals into (between) choruses, etc. It could be controlled/triggered by MIDI and did pitch shifting too, so may well have been used for the "tay-tay-tay-tay" on Mel & Kim's "Respectable."

I have no idea why it was named the "Infernal Machine" (it's French)--but just looking at it hardly inspires confidence in it delivering impeccable performance...

Or maybe the test tone was recorded to a Fairlight II--LOL.

Then again, I might just be talking total nonsense and the tone generator is to blame:

The other one looks like one of those hobby oscillators with a diode shaper for a sine wave and a horrible power supply.

 

[mansr]

Checking a few periods, I counted 44 samples between zero-crossings.

44100/44 = 1002.272727... Hz.

44100 doesn't divide exactly by 1000. The 1000 Hz tone will have 44 samples in most periods and 45 in a few for an average of 44.1. The signal repeats exactly with a period of 441 samples. This results in a main tone of exactly 1000 Hz. The rounding errors repeat at 100 Hz which is why the spectrum has spikes at that interval.

 

[xr100]

44100 doesn't divide exactly by 1000. The 1000 Hz tone will have 44 samples in most periods and 45 in a few for an average of 44.1.

D'oh! Thanks.

OK, here's with a period of 44.1 samples (note the amplitude range is now down to -180dB):

And the Denon:

Slightly better than the Denon but not much difference.

The infamous "DF Metric" result is -38.9dB (as calculated by DeltaWave.)

 

[xr100]

That might not be literally true but many processors have undocumented op-codes. Intel has a list of benchmarks for common math operations using their compiler and processors vs gcc.

Maybe the claim ESS makes of improved Delta-Sigma converter performance are akin to claims of improved CPU pipeline performance that may or may not be true, or are exaggerated, and may or may not be of practical utility.