Article: Understanding Digital Audio Measurements

[guigui]

SINAD in these devices is usually dominated by noise. There, there can be small differences due to ground current/loops, that account for some of that difference. Other differences could be real but is really not material. FYI in my SINAD ranking graph, I use the average of the two channels.

Thanks a lot, Amir!

 

[danhilu]

Hi Daniel. I see that my explanation was perhaps not as clear as it could be. But I got lost half-way through yours as it got too long!

Fair enough.
Short version: the statement "So if I set the digital generator to -120 dBFS, it means I am telling the DAC to reproduce a voltage that is represented by 120/6 = 20 bits." is wrong.
-120 dBFS corresponds to a signal where 20 bits are unset - using @danadam 's phrasing. The voltage is "represented" by at most 4 bits - using your phrasing @amirm.
120 dB "corresponds" indeed to 20 bits, but you have to remove these from the available 24 bits. As I've laboriously tried to explain above

 

[danhilu]

@danhilu Well, this also wrong; because if you want to be accurate, at 16 bits the dynamic range is ~90dB and ~138dB at 24 bits (if we do not dither). This is because audio samples are signed number. Which implies another inaccuracy in your post above: the most significant bit is always “used”.

Thank you @melowman. Excellent. Signal also need to get negative. You're right of course.

 

[danhilu]

I think you are confusing bits needed with bits set/unset. With lower amplitudes the top bits are unset (ignoring two's complement for now) but they are still needed.

Thank you @danadam . I am not convinced by my wording either. I was trying to transform an input statement which I have understood is not correct into something a layman with some motivation could decipher. I think you're right, "unset" is better than "needed". Bottom line is my reply to @amirm above: I think there is an error in the initial text.

 

[LionIT]

What is the type of window that AP applies to make FFT of various measurements? Rectangular? Blackman Harris?
I noticed in REW that the window type changes the results enormously sometimes, especially at very low levels.

 

[rascal101]

While I understand that the measurements pertain mostly to the frequency region, how about the accuracy of the output for the D/A converter.

Let's do some basic calculations to see what are the effects of a 16 bit DAC and a 32 bit DAC.

I would like to point here that although Delta Sigma DAC architecture is GOOD in theory it still succumbs to the limitation of the Input Power Supply and Voltage Reference (Vref) Precision because let us face it a DAC will convert Digital Signal to Analog Signal. And the accuracy of the Analog Signal is a reflection of the quality of the Power Supply and Voltage Reference wherein the Analog Signal is pegged.

I will be using the AKM AD4399 DAC, Burr Brown PCM56 and Analog Devices LT6657 voltage reference and a precision 5V power supply with 0.1% accuracy as references to the calculation.

AKM4399 is a 32 bit Delta Sigma D/A IC
Burr Brown PCM63 is a 16 bit D/A IC
Analog Devices LT6657 is a precision 1.25V reference IC with 2ppm/V Line Regulation and 3ppm/C voltage drift

For the 5V regulator with 0.1% tolerance, Vmin=4.995V and Vmax=5.005V -- this represents 10mV tolerance
For 16 bit, calculating the minimum voltage it can handle, using the LT6657, it is 1.25/2^16=19.1nV
For 32 bit, calculating the minimum voltage it can handle, again using the LT6657, it is 1.25/2^32=0.29nV

Calculating the accuracy of the LT6657,
with the 2ppm/V Line Regulation, and using the 5V 0.1% regulator with the 10mV tolerance, LT6657 will drift by (2/1,000,000) x 0.01 = 20nV
with the 3ppm/C voltage drift over temperature and using 5C drift in typical household use, LT6657 will drift by (3/1,000,000) x 5 = 15,000nV

Total drift of the LT6657 will then be 15,000nV + 20nV = 15,020nV

Looking into this figures, we can see that the drift alone by the LT6657 is so much large than what a 32bit DAC can handle because 15,020nV >> 0.29nV

However, for the 16bit DAC 19.1nV << 15,020nV, it is well within the limitations of the Voltage Reference

The point here is increasing the number of bits and adding more filters will of course result is better measurements. However, let us take note that we are using our ears to "measure" music not instruments. If we would like the precision of the D/A IC, we need to at least have same level of equipment being used in the factory setting wherein D/A IC are being tested/measured.

As I know, measurement instruments in the factory are calibrated precisely and these are using 0.1% accuracy components further the test fixtures used have layout that only commercially produced DAC (completed DAC machine not the D/A IC) are not able to duplicate.

So, to answer the question, will the AKM4399 beat the PCM63

1. In the factory setting, that is a resounding -- YES
2. In the commercial, that is a resounding -- I DO NOT KNOW, it depends on how the D/A was layed out on the PCB and the accuracy of the Power Supply and the
stability of the temperature which the D/A IC operates

From this, I will take on the PCM63 anyday vs the AKM4399. I have both D/A IC based DAC and no matter how I listen to the AKM4399 it is sterile -- too much L-C tank filter. It is like eating food at the hospital -- almost no salt.

 

[rascal101]

I would like to point out that measurements should not emphasize only on time or frequency, it should also be measured vs the actual RMS Voltage or RMS Current that is the OUTPUT of the D/A converter. Resolving time related issues and not looking into magnitude based issues does not complete resolve issues with D/A conversion. One also must look precisely with the Linear Regulator and Voltage Reference. These have tolerances that must be precisely controlled to minimize their effects. And understand their limitations so that arguments into bit count and time related issues can be properly understood.

 

[rascal101]

I think one problem for typical customers unaware of the effects of proper PCB layout techniques specially for analog circuitry is how to distinguish a well layed out DAC and improperly layed out DAC. For me, I will take my time to review how the engineering reputation of the DAC designer/manufacturer. Also to check a reviewer who checks the PCB layout and components used. And check if the DAC design adheres to the application note put out by the D/A IC designer/manufacturer.

 

[rascal101]

While I understand that the measurements pertain mostly to the frequency region, how about the accuracy of the output for the D/A converter.

Let's do some basic calculations to see what are the effects of a 16 bit DAC and a 32 bit DAC.

I would like to point here that although Delta Sigma DAC architecture is GOOD in theory it still succumbs to the limitation of the Input Power Supply and Voltage Reference (Vref) Precision because let us face it a DAC will convert Digital Signal to Analog Signal. And the accuracy of the Analog Signal is a reflection of the quality of the Power Supply and Voltage Reference wherein the Analog Signal is pegged.

I will be using the AKM AD4399 DAC, Burr Brown PCM56 and Analog Devices LT6657 voltage reference and a precision 5V power supply with 0.1% accuracy as references to the calculation.

AKM4399 is a 32 bit Delta Sigma D/A IC
Burr Brown PCM63 is a 16 bit D/A IC
Analog Devices LT6657 is a precision 1.25V reference IC with 2ppm/V Line Regulation and 3ppm/C voltage drift

For the 5V regulator with 0.1% tolerance, Vmin=4.995V and Vmax=5.005V -- this represents 10mV tolerance
For 16 bit, calculating the minimum voltage it can handle, using the LT6657, it is 1.25/2^16=19.1nV
For 32 bit, calculating the minimum voltage it can handle, again using the LT6657, it is 1.25/2^32=0.29nV

Calculating the accuracy of the LT6657,
with the 2ppm/V Line Regulation, and using the 5V 0.1% regulator with the 10mV tolerance, LT6657 will drift by (2/1,000,000) x 0.01 = 20nV
with the 3ppm/C voltage drift over temperature and using 5C drift in typical household use, LT6657 will drift by (3/1,000,000) x 5 = 15,000nV

Total drift of the LT6657 will then be 15,000nV + 20nV = 15,020nV

Looking into this figures, we can see that the drift alone by the LT6657 is so much large than what a 32bit DAC can handle because 15,020nV >> 0.29nV

However, for the 16bit DAC 19.1nV << 15,020nV, it is well within the limitations of the Voltage Reference

The point here is increasing the number of bits and adding more filters will of course result is better measurements. However, let us take note that we are using our ears to "measure" music not instruments. If we would like the precision of the D/A IC, we need to at least have same level of equipment being used in the factory setting wherein D/A IC are being tested/measured.

As I know, measurement instruments in the factory are calibrated precisely and these are using 0.1% accuracy components further the test fixtures used have layout that only commercially produced DAC (completed DAC machine not the D/A IC) are not able to duplicate.

So, to answer the question, will the AKM4399 beat the PCM63

1. In the factory setting, that is a resounding -- YES
2. In the commercial, that is a resounding -- I DO NOT KNOW, it depends on how the D/A was layed out on the PCB and the accuracy of the Power Supply and the
stability of the temperature which the D/A IC operates

From this, I will take on the PCM63 anyday vs the AKM4399. I have both D/A IC based DAC and no matter how I listen to the AKM4399 it is sterile -- too much L-C tank filter. It is like eating food at the hospital -- almost no salt.

Note: above calculation also need to consider Load Regulation for the Voltage Reference. So, this means the accuracy of the LT6657 is much actually worse.

This is not to say the LT6657 is bad rather, it is limited. It is already a very good reference. The question is how good is the internal Voltage Reference of the AKM4399? Or the PCM63?

Who among the various DAC designer/manufacturer have good reputation for designing very good voltage references? Does AKM have better in house design than TI (acquired Burr Brown)? How about ESS or Crystal Semiconductor?

 

[amirm]

AKM4399 is a 32 bit Delta Sigma D/A IC

Nothing is 32 bits. Above only means 32 bit samples are accepted, not that this is properly output. We don't even have a "24 bit" DAC. Best case has been 22 bits (measured).

 

[rascal101]

Nothing is 32 bits. Above only means 32 bit samples are accepted, not that this is properly output. We don't even have a "24 bit" DAC. Best case has been 22 bits (measured).

I understand that. That is why I did some calculations. Not even in actual design and setting. In theory yes.

 

[Blumlein 88]

@rascal101
Welcome to ASR.

The measurements in the reviews here are the actual analog output of DACs. So one can conjecture about how well something is done, and certainly the components may limit performance. The point is for commercial products measured here you can see what results are. Some devices come quite close to the reference designs and some fall woefully short. Measurements are done with one of the best possible test devices available.

As far as how one Dac sounds vs another you run into this quite often on the forum. Did you match levels precisely while comparing? As in .1 db or less difference. If not, the comparison as you heard it is not reliable. It also is better if you had someone do switching so the comparison was blind. Every human is easily biased by many things to confound such results.

 

[rascal101]

Yes you are correct. The listening is subjective however, the measurements seen only show 1 part and focused mostly on time and frequency -- SNR vs frequency. My point here is there should be measurement on DNL, INL etc not just frequency response. That is why I showed calculations relative to the input power supply and voltage reference used by DAC. How good are they? You should also measure the accuracy of the output rms voltage or current vs expected output to check for linearity. Previous posts do not clearly indicate this. The point here is even with 16 bits it is already overkill because most systems do not have resolution to accomplish this. Unless you have a reference design and you are using factory correlated test fixtures and instruments.

 

[Blumlein 88]

Yes you are correct. The listening is subjective however, the measurements seen only show 1 part and focused mostly on time and frequency -- SNR vs frequency. My point here is there should be measurement on DNL, INL etc not just frequency response. That is why I showed calculations relative to the input power supply and voltage reference used by DAC. How good are they? You should also measure the accuracy of the output rms voltage or current vs expected output to check for linearity. Previous posts do not clearly indicate this. The point here is even with 16 bits it is already overkill because most systems do not have resolution to accomplish this. Unless you have a reference design and you are using factory correlated test fixtures and instruments.

Amir shows linearity in every review of DACs. I've looked into the DNL of a few DACs. The good sigma delta DACs appear to put out nigh on perfect results until the signal gets low enough to be swamped by noise. It is a solved issue. No need to keep proving the round wheel rolls.

BTW, if you wanted to do your tests, there is some good software not available when I last looked into it which would make it easier.
A member here pkane has written some useful tools. One of which is Multitone which has grown into a test suite of many things.
Thread here:

Beta Test: Multitone Loopback Analyzer software

This is a software beta-test for IMD and THD aficionados with an intense desire for multi-tone audio testing :) The start of the documentation and the download website is here: https://distortaudio.org/multitone.html (for 64-bit Windows) Multi-tone Analyzer (MA) was more of an experiment than...

www.audiosciencereview.com

Download it here:

Multitone - Loopback Analyzer for Audio | Distort documentation

Multitone is currently in beta test. Use at your own risk!

distortaudio.org

Another is Deltawave which is a digital null testing software. Among other things it will show linearity using music for the test signal.

Beta-test: DeltaWave Null Comparison software

Folks, Only for the brave: I'd like to solicit your feedback and suggestions on DeltaWave null-testing software. It's an early preview, so please be gentle! Details and software download are available here: https://deltaw.org An example report produced by DeltaWave comparing two USB cables...

www.audiosciencereview.com

You can download it from here:

DeltaWave Audio Null Comparator | DeltaWave documentation

deltaw.org

Feed in the digital original file, and compare with the output of a DAC and you can see the linearity.

 

[rascal101]

I am just writing because the R2R ladder is not being given the response it deserves. What some people do not realize is it is far faster to set the correct output on R2R ladder vs Delta Sigma. The Delta Sigma requires a feedback loop to correct itself. From this we can see the response for Delta Sigma is "slow" vs the R2R ladder. If one can create very precise resistors, R2R will beat Delta Sigma hands down.

 

[Blumlein 88]

I am just writing because the R2R ladder is not being given the response it deserves. What some people do not realize is it is far faster to set the correct output on R2R ladder vs Delta Sigma. The Delta Sigma requires a feedback loop to correct itself. From this we can see the response for Delta Sigma is "slow" vs the R2R ladder. If one can create very precise resistors, R2R will beat Delta Sigma hands down.

If one can create very precise resistors.....

And there in lies the rub. What evidence do you have that R2R is better at linearity than delta sigma? We cannot create resistors precise enough for what you propose. 20 to 24 bit linearity?

 

[rascal101]

Also, anything with feedback loop is subject to instability.

 

[Blumlein 88]

Also, anything with feedback loop is subject to instability.

Super precise resistors in R2R DACs are subject to instability too. Temperature for one. And the power supply as you've noted for another. How do you explain the results of sigma delta if they are typically suffering from instability. While it could happen, the makers have worked out how to make it work without that being a problem.

 

[rascal101]

If one can create very precise resistors.....

And there in lies the rub. What evidence do you have that R2R is better at linearity than delta sigma? We cannot create resistors precise enough for what you propose. 20 to 24 bit linearity?

Unfortunately, that is the limitation. But the R2R ladder for me is "better". Too much math for the Delta Sigma for me ... hahaha ... makes my head ache. Further, designing good loop response ... that is the challenge ...

 

[rascal101]

Super precise resistors in R2R DACs are subject to instability too. Temperature for one. And the power supply as you've noted for another. How do you explain the results of sigma delta if they are typically suffering from instability. While it could happen, the makers have worked out how to make it work without that being a problem.

Feedback loop for correction ...