DAC/amp performance with true peaks over 0 dBFS

[kemmler3D]

A very significant proportion of the CDs I checked so far have over-0 true peak. Maybe half of them? Here are some more examples if anyone is interested. Many of these are from the 1980s and not very compressed. There are a couple of outliers: Billie Jean on Thriller +3.19 dBFS, Valotte on Music Life Virgin Version +3.04 dBFS.

Heartbeat City (The Cars) https://www.discogs.com/release/13924732-The-Cars-Heartbeat-City
It's Not The Night +0.57 dBFS (track DR 14)

Thriller (Michael Jackson) https://www.discogs.com/release/1782800-Michael-Jackson-Thriller
Billie Jean +3.19 dBFS (track DR 14)

Storm Front (Billy Joel) https://www.discogs.com/release/6304490-Billy-Joel-Storm-Front
Leningrad +0.03 dBFS (track DR 14)

In The Digital Mood (The Glenn Miller Orchestra) https://www.discogs.com/release/34371247-The-Glenn-Miller-Orchestra-In-The-Digital-Mood
+0.89 dBFS (track DR 14)

Legend (Clannad) https://www.discogs.com/release/5129078-Clannad-Legend
Battles +1.57 dBFS (track DR 14)

Faith (George Michael) https://www.discogs.com/release/28682467-George-Michael-Faith
+0.91 dBFS (track DR 13)

Whenever You Need Somebody (Rick Astley) https://www.discogs.com/release/1513322-Rick-Astley-Whenever-You-Need-Somebody
+0.11 dBFS (track DR 14)

Music Life - Virgin Version https://www.discogs.com/release/1546667-Various-Music-Life-Virgin-Version
Valotte +3.04 dBFS (track DR 14)

Revenge (Eurythmics)
The Last Time +0.28 dBFS (track DR 14)

...But Seriously (Phil Collins) https://www.discogs.com/release/13712405-Phil-Collins-But-Seriously
Heat On The Street +0.59 dBFS (track DR 14)

No Jacket Required (Phil Collins) https://www.discogs.com/release/13748012-Phil-Collins-No-Jacket-Required
Sussudio +0.47 dBFS (track DR 16)

Face Value (Phil Collins) 16029-2 / 299 143 https://www.discogs.com/release/2049760-Phil-Collins-Face-Value
Highest true peak Behind The Lines -0.45 dBFS (track DR 15)
(So no tracks over 0 dBFS.)

Face Value (Phil Collins) CDV 2185 https://www.discogs.com/release/10058743-Phil-Collins-Face-Value
I'm Not Moving +0.46 dBFS (track DR 15)
Total 7 tracks over 0 dBFS, though four only very slightly: +0.02 to +0.08.

Face Value (Phil Collins) 20P2-2074 https://www.discogs.com/release/4030076-Phil-Collins-Face-Value
Behind The Lines +0.35 dBFS (track DR 15)
Total 5 tracks have true peak over 0 dBFS.

With lots of real world examples I guess the thing to do now is look for audibility of defects among them...

 

[jkim]

I believe, going forward, ANYONE wanting to discuss issues related to ISO/ISP should first have a look at Archimago's recent study on the topic. As usual, it is an amazing, comprehensive analysis!

And I agree to his conclusions:

My recommendation of +3dB discussed a few years back for DAC intersample overhead I think remains fair. This will easily cover the vast majority of music out there (including 96% of my Electronica tracks). However, if you're very much into electronica/techno/EDM, maybe a DAC with +4.25dB headroom would given you a bit more assurance - or simply attenuate your DAC digital volume by a few dBs, or even better, use ReplayGain with something like -18 LUFS target.

And more importantly,

I think the ultimate conclusion to all of this is simply this: instead of insisting that our DACs have high intersample overload headroom, it would have been much better if the music were created respecting 0dBFS as the limit. I would argue that music with high intersample peaks - let's say more than +1dB which is higher than 99% of my classical music collection is likely not great sounding "hi-fi" material anyways.

I would still like to have 3 dB headroom for peace of mind. IMO the best, device-side solution to ISO is software volume control (in DSP before signal sent to a DAC chip), like in RME ADI-2 series or Qudelix 5K.

 

[JSmith]

How many threads are there on this topic?

Some consolidation would be good...


JSmith

 

[CuriousAli]

Sorry for asking the obvious question, I use both ReplayGain and have set the Master Fader of my audio interface to -4dB, so am I safe regarding this for listening to music?

 

[oleg87]

Sorry for asking the obvious question, I use both ReplayGain and have set the Master Fader of my audio interface to -4dB, so am I safe regarding this for listening to music?

View attachment 489021

If the signal you're feeding to your DAC is already comfortably away from 0dBFS then yeah, not much chance of ISOs.

 

[CuriousAli]

If the signal you're feeding to your DAC is already comfortably away from 0dBFS then yeah, not much chance of ISOs.

Thanks, I'm using the DAC over USB and playing tracks with ReplayGain applied to them so they are well below that I think.

 

[Donuts123]

I'm not familiar with how ReplayGain works. If a particular track is adjusted so its sample peak is close to 0 dBFS, there could still be issues depending on your your DAC operates. E.g. if it resamples/oversamples before adjusting gain.

 

[jkim]

I'm not familiar with how ReplayGain works. If a particular track is adjusted so its sample peak is close to 0 dBFS, there could still be issues depending on your your DAC operates. E.g. if it resamples/oversamples before adjusting gain.

ReplayGain in foobar2000 has a "true peak scan" function with oversampling (x2, x4, x8) options. There are other add-on components for more advanced configuration (target LUFS, maximum true peak, etc.), too.

 

[OHtaru]

I think we've ignored one thing: Can true peak scan really represent what happens in the DAC?

If the DAC uses a linear phase filter, it's likely yes. If the DAC uses a minimum phase filter, then no.

Let's see how the minimum phase filter increases the peak level. This is one of the reasons why I prefer linear phase filters.
If you prefer minimum phase filter, you may even need more headroom than Archimago's study. As for how much is needed, I think research in this area is still a blank.

 

[DVDdoug]

I'm not familiar with how ReplayGain works. If a particular track is adjusted so its sample peak is close to 0 dBFS, there could still be issues depending on your your DAC operates. E.g. if it resamples/oversamples before adjusting gain.

The file is scanned for loudness and peak level.

"True peak" scanning is optional and not always implemented. (I don't believe it's part of the ReplayGain standard.)

There is also more than one way to calculate loudness. ReplayGain was created before the EBU R128/LUFS standard so sometimes it uses the LUFS and sometimes it uses its own algorithm.

The target loudness is set so that MOST tracks have to be lowered in volume. That's because many quiet-sounding tracks have 0dB peaks and they can't be boosted. The only way to volume match (without clipping) is to lower the volume of most tracks.

Since most tracks are reduced in volume most tracks won't have any inter-sample overs.

But SOME tracks are boosted and in that case ReplayGain could CAUSE inter-sample overs (if your implementation isn't checking the "true peaks").

Even with the low target loudness, some quiet-sounding tracks still need to be boosted to hit that target. Normally ReplayGain will only boost until the new peaks are 0dB so you don't get clipping. So there are some tracks that are still quieter than others. Usually it's jus checking the raw data peaks but some implications can keep the "true peaks" 0dB or less. (There is an option to ignore the peaks and allow clipping.)

I think we've ignored one thing: Can true peak scan really represent what happens in the DAC?

Right. There is not "one true peak". It depends on the DAC so it's just an approximation unless you are measuring the actual analog output from the DAC.

 

[jkim]

I think we've ignored one thing: Can true peak scan really represent what happens in the DAC?

If the DAC uses a linear phase filter, it's likely yes. If the DAC uses a minimum phase filter, then no.
View attachment 489918
Let's see how the minimum phase filter increases the peak level. This is one of the reasons why I prefer linear phase filters.
If you prefer minimum phase filter, you may even need more headroom than Archimago's study. As for how much is needed, I think research in this area is still a blank.

True peaks in the audio band?

 

[Donuts123]

True peaks in the audio band?

If the fitered signal clips then the distortion introduced would have components in the audio band I think.

 

[OHtaru]

If the fitered signal clips then the distortion introduced would have components in the audio band I think.

Indeed.

I did a null test using upsampled Nakashima Mika - Be Real 32-bit float and 16-bit int (clipped the ISPs) versions, and the result is that the distortion falls within the entire audible frequency range:

The general method of ISP tests always appear as ultrasonic harmonics.
However, the null test is one thing, and audibility is another. I ABXed the 32-bit float and 16-bit int versions, and I couldn't distinguish between them.

 

[jkim]

Indeed.

I did a null test using upsampled Nakashima Mika - Be Real 32-bit float and 16-bit int (clipped the ISPs) versions, and the result is that the distortion falls within the entire audible frequency range:View attachment 489951
The general method of ISP tests always appear as ultrasonic harmonics.
However, the null test is one thing, and audibility is another. I ABXed the 32-bit float and 16-bit int versions, and I couldn't distinguish between them.

Can you give more specifics on how this test is done? How did you create the two versions? And what is the FFT setting? I am trying to understand how digital clipping due to an intersample peak at a high frequency can propagate into lower frequencies. I can understand it if the phenomenon is from an analog amplifier. But this is purely in the digital domain.

 

[OHtaru]

Can you give more specifics on how this test is done? How did you create the two versions? And what is the FFT setting? I am trying to understand how digital clipping due to an intersample peak at a high frequency can propagate into lower frequencies. I can understand it if the phenomenon is from an analog amplifier. But this is purely in the digital domain.

Version A：Upsampled 8x 44.1k to 352.8k, saved as 32 bit float file.
Version B：Upsampled 8x 44.1k to 352.8k, saved as 16 bit int file (ISP was clipped).
Null test: Version A minus Version B.
You can see that the remaining part consists of some pulses.

As for why the real music ISP exhibits full-frequency distortion and the 11025Hz ISP exhibits ultrasonic harmonics, it is because the real music ISP consists of random pulses, while the 11025Hz ISP consists of periodic pulses. The pulse itself is full-frequency. Once the pulses are arranged periodically, within a certain time window (FFT size), they will appear as a series of harmonics of the periodic frequency. This is that in the Fourier transform, periodicity in the time domain corresponds to discreteness in the frequency domain, and aperiodicity in the time domain corresponds to continuity in the frequency domain.

 

[jkim]

Version A：Upsampled 8x 44.1k to 352.8k, saved as 32 bit float file.
Version B：Upsampled 8x 44.1k to 352.8k, saved as 16 bit int file (ISP was clipped).
Null test: Version A minus Version B.
You can see that the remaining part consists of some pulses.
View attachment 490097View attachment 490098
As for why the real music ISP exhibits full-frequency distortion and the 11025Hz ISP exhibits ultrasonic harmonics, it is because the real music ISP consists of random pulses, while the 11025Hz ISP consists of periodic pulses. The pulse itself is full-frequency. Once the pulses are arranged periodically, within a certain time window (FFT size), they will appear as a series of harmonics of the periodic frequency. This is that in the Fourier transform, periodicity in the time domain corresponds to discreteness in the frequency domain, and aperiodicity in the time domain corresponds to continuity in the frequency domain.

View attachment 490103View attachment 490102

Thank you for taking time to describe this. Enlightening. In fact, if you used a smaller FFT size, the effect would have been even greater! With 4096, the effect is highly diluted.