Ending the Windows Audio Quality Debate

[Roen]

Don't you switch between outputs manually in PEACE? How does APO know?

You see the dropdown on top of the editor that says Device?

If you set it up this way, you can see the EQ curve on the bottom change as you change device.

However, if you unplug a headphone and swap to another headphone, you'd have to manually change the EQ curve in EAPO at the DAC device level. This works if you have a DAC, another output dedicated for another device and BT headphones let's say.

 

[Roen]

Does the -4 dB to avoid inter-sample overs setting still apply when using ASIO / WASAPI (i.e. foobar2000 with MathAudio Headphones EQ)?

 

[Davide]

Does the -4 dB to avoid inter-sample overs setting still apply when using ASIO / WASAPI (i.e. foobar2000 with MathAudio Headphones EQ)?

Yes. The problem does not depend on the interface drivers, but on the conversion in dac ic.

 

[Roen]

Yes. The problem does not depend on the interface drivers, but on the conversion in dac ic.

In foobar, should I apply -4.0 dB volume reduction or -4.0 dB MathAudio Headphones EQ preamp gain?

 

[Davide]

In foobar, should I apply -4.0 dB volume reduction or -4.0 dB MathAudio Headphones EQ preamp gain?

There shouldn't be any difference

 

[Mulder]

I decided to investigate this as I kept hearing issues with Windows, even after following common practice to (controversially) supposedly make it clean. I found some surprises along the way.

Thank you for sharing your research on this. It was very helpful.

 

[Dunring]

I just came across a Github with a program for disabling the peak limiter. It took Google Chrome translate, but it's got the source code and both 32 and 64 bit versions at:

GitHub - skawamoto0/dpeaklim: Disable Peak Limiter in Windows Audio Engine

Disable Peak Limiter in Windows Audio Engine. Contribute to skawamoto0/dpeaklim development by creating an account on GitHub.

github.com

It was posted three years ago, so the problem might have been fixed with a patch in the meantime. I ran the question through Chatgpt just for kicks, and it reported that the limiter was patched in some versions of Windows 10. It didn't say which ones (later ones I assume).

All it really had to say was:
"The CAudioLimiter issue was a bug in the Windows 10 audio stack that caused distortion and other audio quality problems when certain conditions were met. Microsoft released several updates to address this issue in Windows 10, and it is possible that these updates also apply to Windows 11."

 

[edechamps]

I just came across a Github with a program for disabling the peak limiter. It took Google Chrome translate, but it's got the source code and both 32 and 64 bit versions at:

GitHub - skawamoto0/dpeaklim: Disable Peak Limiter in Windows Audio Engine

Disable Peak Limiter in Windows Audio Engine. Contribute to skawamoto0/dpeaklim development by creating an account on GitHub.

github.com

In case you haven't seen, I already found a way to do that using a small registry hack. Which is a hack yes, but is still vastly cleaner than what the program you linked is doing: that program is monkey-patching the audio engine binary which is very risky and could potentially cause all kinds of problems, especially if you run it on a Windows version it was not designed for.

 

[meleeli]

Have tried to follow this guide and subsequent discussion, but I'm unsure about a few things.

1. Does it matter if the preamp is part of eqapo's pre-mix or post-mix?

2. Following on from that. Let's say I have a post-mix EQ profile with a peak gain of -1db. Should I have the pre-mix preamp be -3db so the total peak gain is -4db, or should the preamp be -4db on its own?

3. If it's not relevant to define pre/post-mix stages, is it safe to assume the peak gain would just need to total -4db?

 

[edechamps]

Have tried to follow this guide and subsequent discussion, but I'm unsure about a few things.

1. Does it matter if the preamp is part of eqapo's pre-mix or post-mix?

No. It's floating point all the way through anyway. All that matters is the final result post mix.

2. Following on from that. Let's say I have a post-mix EQ profile with a peak gain of -1db. Should I have the pre-mix preamp be -3db so the total peak gain is -4db, or should the preamp be -4db on its own?

It's the total gain that counts.

3. If it's not relevant to define pre/post-mix stages, is it safe to assume the peak gain would just need to total -4db?

Yes.

Note that, in general, it's preferable to use post-mix filters rather than pre-mix filters because post-mix filters are more efficient (they only run once, instead of once per stream). For most use cases there is no reason to use pre-mix filters.

 

[Davide]

I wanted to carry out more in-depth tests of the Win audio system.

I have Win 11 latest version, with Virtual Cable HiFi ASIO Bridge installed to make loopback measurements with REW.

Since I'm tormented by the fact that the Win mixer doesn't automatically switch the frequency, I wanted to include the resampling in the test, i.e. from 44.1/16 to 192/24, representative of normal playback with Spotify for example. So I set Virtual Cable to 192kHz both in and out.

In REW I set 44.1kHz with default JAVA driver (so no Exclusive Mode) and then played three types of signals (16 bit dithered) to make sure I don't see just THD+N but also TD+N, which I feel hasn't been checked in this thread.

For TD+N I used multitone and triple tone (10.5 + 19 + 20 kHz). However, the signal parameters are indicated in the following screens.

TO NOTE for multitones are the peak dBFS of signal, i.e. those indicated in small print at the bottom right of the RMS signal level box.

Also note the peak dBFS of captured signal in top right box, as well as the distortion in the upper left box on the graphs.

The interesting thing is that with triple tone -0.1 dBFS pk is enough to avoid any distortion, while with single tone -0.2 dBFS pk is needed.
With multitone instead you need -2.77 dBFS pk. Above that value the TD distortion reaches 0.75%. Unacceptable.

I'm not an expert on signals but it seems to me that with complex signals at least -3 dBFS peak is needed to avoid the Windows limiter. The single tone test is not sufficient to show the intervention threshold.

What is not clear to me is on the basis of which level the limiter intervenes.

However, the positive thing is that without the limiter the resampling from 44.1/16 to 192/24, that always occurs in these test, does not involve any audible distortion.
To evaluate the goodness of the resampler it would be necessary to do other tests, I know. But for people like me who are concerned about the fact that normal streaming music players don't support really bit perfect drivers (see WASAPI Excl. mode of Amazon Music) know we can upsample everything to 192khz without audible distortion is a nice relief.
For the record, I also made measurements with 192kHz signals in order to avoid resampling, and it changes about 0.005% THD in favor (worst case with multitone signal).

If I have done something wrong, or there are any distortions not detected/detectable by REW RTA please let me know, thanks.

Here the results:

 

[Davide]

I add two more interesting measurements.
The first with 1kHz square wave at 0 dBFS peak (0 dBFS RMS), again with 44.1/16 --> 192/24 resampling. A lot of additional signals appear, but I don't know how significant it is honestly.

And then the full band sweep measurement always 0 dBFS peak (-3,01 dBFS RMS). Note the different distortion level as a function of frequency.

 

[BeerBear]

The interesting thing is that with triple tone -0.1 dBFS pk is enough to avoid any distortion, while with single tone -0.2 dBFS pk is needed.
With multitone instead you need -2.77 dBFS pk. Above that value the TD distortion reaches 0.75%. Unacceptable.

I'm not an expert on signals but it seems to me that with complex signals at least -3 dBFS peak is needed to avoid the Windows limiter. The single tone test is not sufficient to show the intervention threshold.

What is not clear to me is on the basis of which level the limiter intervenes.

I checked out that REW multitone signal. When Windows resamples it from 44.1k to 192k, the sample peak increases by 2.64dB. That's why I think you're seeing the limiter kick in already at -2.77dB at the input (-2.77dB would be -0.13dB after resampling). So, it looks like Windows resamples first and then looks at the sample peak level to decide whether to limit or not.

Note that even a high quality SRC is likely to increase the sample peak level for such a signal. I tried resampling it with SoX and it was the same.
And if you play it at 44.1k without resampling (maybe with ASIO or some exclusive mode), it's likely to have similar peak increases during the oversampling that happens inside the DAC, which could result in audible clipping.

 

[Davide]

Yeah, intersample overs. The problem is when they exceed 0 dBFS due tu upsampling (of bad source signal, badly limited).
However, it can be deduced that the Win limiter does not prevent true peaks that take place inside the DAC, as professional plugins do.
It will be up to the manufacturer of the DAC to manage the intersamples.
Meanwhile on the Win side it is sufficient to put an attenuation of -3.5 dB with Eq APO to solve the problem.
In fact I tried to put this attenuation and reproduce a triple tone at 0 dBFS peak, always with upsampling 44,2/16 to 192/24, and the result is this:

For peace of mind I also tried resampling from 48, 88.2, 96 formats and nothing changed..
As far as I'm concerned this result is enough to sleep peacefully. I'll leave the Win mixer fixed at 192/24 so I don't have to worry about the source format anymore.

I only have the doubt that at this point to prevent intersample overs inside the DAC (where not managed like Benchmark and RME for example) a further -3.5 dB of attenuation is needed.

 

[Robbo99999]

Yeah, intersample overs. The problem is when they exceed 0 dBFS due tu upsampling (of bad source signal, badly limited).
However, it can be deduced that the Win limiter does not prevent true peaks that take place inside the DAC, as professional plugins do.
It will be up to the manufacturer of the DAC to manage the intersamples.
Meanwhile on the Win side it is sufficient to put an attenuation of -3.5 dB with Eq APO to solve the problem.
In fact I tried to put this attenuation and reproduce a triple tone at 0 dBFS peak, always with upsampling 44,2/16 to 192/24, and the result is this:
View attachment 298020

For peace of mind I also tried resampling from 48, 88.2, 96 formats and nothing changed..
As far as I'm concerned this result is enough to sleep peacefully. I'll leave the Win mixer fixed at 192/24 so I don't have to worry about the source format anymore.

I only have the doubt that at this point to prevent intersample overs inside the DAC (where not managed like Benchmark and RME for example) a further -3.5 dB of attenuation is needed.

Regarding intersample overs, when I've measured a lot of my tracks using Orban Loudness Meter, I worked out that using a negative preamp of -2dB will cut out almost all of them. (I don't really want to use more than that when combining it with a negative preamp that is also associated with EQ boosts, so I figure -2dB on top of whatever your EQ requires is gonna be sufficient because the negative preamp of the EQ is also gonna cover off a fair degree of the intersample overs anyway, so the additional -2dB is just to be sure.)

 

[Davide]

Regarding intersample overs, when I've measured a lot of my tracks using Orban Loudness Meter, I worked out that using a negative preamp of -2dB will cut out almost all of them. (I don't really want to use more than that when combining it with a negative preamp that is also associated with EQ boosts, so I figure -2dB on top of whatever your EQ requires is gonna be sufficient because the negative preamp of the EQ is also gonna cover off a fair degree of the intersample overs anyway, so the additional -2dB is just to be sure.)

The worst case would happen with a signal around 11kHz sampled at 44.1kHz, where peaks can reach 1.414 cover factor = 3.01dB.
Signal sampled at higher freq shouldn't have any problems because the peaks move in the ultrasonic band.
The speech of the EQ I did not understand what it has to do with it... it can raise or lower the level at different frequencies, depending on how used.

 

[Robbo99999]

The worst case would happen with a signal around 11kHz sampled at 44.1kHz, where peaks can reach 1.414 cover factor = 3.01dB.
Signal sampled at higher freq shouldn't have any problems because the peaks move in the ultrasonic band.
The speech of the EQ I did not understand what it has to do with it... it can raise or lower the level at different frequencies, depending on how used.

Ok, I don't understand how you worked out the 11kHz, but I pretty much totally believe you, so that's fine. Regarding the EQ, when you EQ headphones most of the boost is going into the bass, so you might have up to +9dB bass boost at 20-30Hz which would mean you'd put in at least -9dB Negative Preamp to cover that off to prevent digital clipping. Given that this means that only 20-30Hz is close to 0dBFS then this means areas like 1kHz are likely to be around -9dBFS, so depending on where your intersample overs are taking place (I don't know where they do actually take place but you say 11kHz) then in this example you could have around -9dB already covering off intersample overs in some areas of the frequency response (most normally most areas over 100Hz, due to the biggest EQ boost happening in the low bass).

 

[Davide]

Ok, I don't understand how you worked out the 11kHz, but I pretty much totally believe you, so that's fine. Regarding the EQ, when you EQ headphones most of the boost is going into the bass, so you might have up to +9dB bass boost at 20-30Hz which would mean you'd put in at least -9dB Negative Preamp to cover that off to prevent digital clipping. Given that this means that only 20-30Hz is close to 0dBFS then this means areas like 1kHz are likely to be around -9dBFS, so depending on where your intersample overs are taking place (I don't know where they do actually take place but you say 11kHz) then in this example you could have around -9dB already covering off intersample overs in some areas of the frequency response (most normally most areas over 100Hz, due to the biggest EQ boost happening in the low bass).

Ok, it make sense now.
Of course, one could make an argument in terms of frequency.
Anyway, I didn't calculate those -3.01dB of worst case. They are indicated by Benchmark.
I did a search to understand if they are representative of the worst case and I found an interesting post where they analyze about 5k music tracks of every genre and format. The result is worrying.
Hydrogen Audio Forum

Also, in the thread they talk about an even worse theoretical case that can reach 11dB intersample peak with resampled square wave.
So that small percentage with peaks over 9 dBFS is plausible.

 

[Sokel]

Ok, it make sense now.
Of course, one could make an argument in terms of frequency.
Anyway, I didn't calculate those -3.01dB of worst case. They are indicated by Benchmark.
I did a search to understand if they are representative of the worst case and I found an interesting post where they analyze about 5k music tracks of every genre and format. The result is worrying.
Hydrogen Audio Forum

Also, in the thread they talk about an even worse theoretical case that can reach 11dB intersample peak with resampled square wave.
So that small percentage with peaks over 9 dBFS is plausible.

For ISO's we use:

11025Hz at 45deg +3db gain or,
7350Hz at 60deg +1.35db gain or,
5512.5Hz at 67.5deg +0.69db gain

Edit:All at 44.1Khz sample rate.

 

[srrxr71]

I don’t know what to say about this. In addition I’ve been having pops and clicks in windows - very dependent on the cable length. It requires a bunch of tweaks to fix.

I just switched to Mac to be safe. I don’t like the interface at all but these issues are just annoying.

In windows I suppose I would have to drop the signal -3dB. The volume control keeps resetting to 100%. Real headache.