True Peak may be a matter of definition...

[danadam]

Nightwish - Highest Hopes (Limited Edition):

auto
x2  1.092665    +0.77 dBTP
x4  1.114916    +0.94 dBTP
x8  1.121315    +0.99 dBTP

SoX Normal
x2  1.370465    +2.74 dBTP
x4  1.370464    +2.74 dBTP

The track with highest true peak was "Sacrament of Wilderness" when using "auto" and "Deep Silent Complete" when using SoX.

I also checked the "Deep Silent Complete" with SoX command line:

]$ sox 11_DeepSilentComplete.flac -n gain -10 rate 88200 stats
             Overall     Left      Right
Pk lev dB      -7.26     -7.26     -9.90

I did x2, x4, x8, x16 and it's the same +2.74 dBTP.

Lastly I used the filter coefficients (in attachment) from ITU-R BS.1770-5 - Algorithms to measure audio programme loudness and true-peak audio level (PDF):

]$ sox 11_DeepSilentComplete.flac -n gain -10 upsample 4 fir bs1770.txt stats
             Overall     Left      Right
Pk lev dB      -7.47     -7.47     -9.84

and it is +2.53 dBTP.

EDIT: And as a bonus, using two different filters in ffmpeg:

]$ ffmpeg -i 11_DeepSilentComplete.flac -af ebur128=peak=true -f null - 2>&1
...
  True peak:
    Peak:        2.6 dBFS

]$ ffmpeg -i 11_DeepSilentComplete.flac -af loudnorm=print_format=json -f null - 2>&1
...
        "input_tp" : "2.63",

 

[AnalogSteph]

How would I do SoX with Upsample x4?

 

[digitalfrost]

View attachment 402160

 

[pLudio]

You should probably have a look at the True Peak Scanner component for fb2k which can also produce several other values like LRA, DR, RMS, LUFS if you enable it in Advanced Settings.

 

[pkane]

https://lufs.org offers an on-line tool (files are not uploaded to the website) for measuring LUFS , TP, and other values from any FLAC or WAV file. They also calculate what will become of your music if you stream it through one of the common streaming services.

 

[AnalogSteph]

View attachment 402202

Bummer, "just" 1.99065 peak amplitude, not even a full +6 dB.

You should probably have a look at the True Peak Scanner component for fb2k which can also produce several other values like LRA, DR, RMS, LUFS if you enable it in Advanced Settings.

That's a good tip, thanks.

Results using that seem to track my previous results with the SoX resampler DSP (upsample 4x, linear phase) fairly closely, with differences that are quite negligible a lot of the time.

Sneaky Sound System:

(vs. 1.608418 - makes sense since min sample rate is set to 192k >176.4k, though if I set SoX to 192k I'm getting a 1.600337 peak; 352.8k gives me 1.615597 at 1338x speed)

Art Angels:

(note the exactly +10.00 LU album LRA, I don't think that's an accident)

Ladyhawke:

Kingdom:

No Hero:

Most of these are identical in the first 3 digits after the decimal point.

So it looks like RG scanning with TP enabled using the SoX DSP with default settings in "Upsample x4" mode makes a pretty good to great approximation for "official" TP scanning while being substantially faster. (I can kinda fudge it by using a slightly wider passband setting, say 95.3%.) Foobar's default "Auto 4x oversample", by contrast, is way off.

As we've previously seen, using a mixed phase filter response increases peak levels. When going all the way to minimum phase, bad ones tend to do so rather dramatically, even more with Quality = "Best". Santogold now becomes this:

...and Sneaky Sound System yields this result:

 

[AnalogSteph]

PS: In order to bring True Peak Scanner RMS output values in line with e.g. the DR Meter plugin, tick the "Use AES +3dB mode" option. Otherwise it will consider a full-scale sinewave to be at -3 dBFS. Yes, that means that there could be signals with higher RMS than peak levels, like a square wave (approximation).

 

[John_Siau]

Several contributors have provided examples of CDs and tracks that show extreme examples of true peak values that are well over 0 dBFS and, in many cases, over +3 dBFS. Thank you for this information.

Are any of you able to provide statistics for your entire music library, or point me to such a database? I need to know what we should expect from the vast majority of CDs and not just the extreme cases cited in the posts above.

As recorded on the CD and analyzed at 4X with a linear-phase filter:

1. What percentage are < 0 dBFS?
2. What percentage are < +1 dBFS?
3. What percentage are < +2 dBFS?
4. What percentage are < +3 dBFS?
5. and so on ...

It is also important to know how many +0 dBFS events occur per minute, if this information is available.

This information would be very helpful to those of us who are designing DACs and other DSP devices.

We need to determine:

1. How much headroom is required to accommodate most CDs?
2. How much headroom is required to accommodate the worst-case CDs?

My standard answer to the first question is that 1 dB is sufficient for 90% of CDs and that 3 dB is enough for 99% of CDs and that you may not care to listen to the last 1%, but I don't have hard numbers to back this up.

 

[AnalogSteph]

A user on HA did that sort of thing a decade ago:

Pathological example of a intersample peak, 11dB

Pathological example of a intersample peak, 11dB

hydrogenaud.io

Although it appears that there were some issues with the data.

My collections has to be about twice the size of theirs, but unfortunately collecting over 15+ years means RG scanning has been carried out with two different algorithms and without / with TP scanning, so it's a bit of a mess that has to be rechecked manually.

Subjectively, I'd say (entire CD TP level):
Moderate overs around 1.05 - like looking for sand on the beach
1.10-1.20 - very common
1.25ish - still fairly common
high 1.30s - seen occasionally but still more common than you'd think
Beyond 1.4, things are rapidly thinning out.
I doubt there are very many releases making to to almost 2.0 like the example posted by @digitalfrost.

I probably have more CDs with true peaks in the high 1.30s and up than emphasis CDs at this point. (Now there's a useless statistic if there ever was one. And no, there is no crossover between the two groups, now that would be an absolute abomination.) It's always a pleasant surprise to find modernish releases with a <0 dBFS TP, they are absolutely out there but a bit of a minority at this point. Maybe 20% tops?

I'd have my doubts about +1 dB doing it for 95% of CDs, but +3 dB for 99% could be about right.

Did some more checking and uncovered two more CDs with TPs of 1.38 to 1.40something, respectively. (Well, make that another three in the 1.40s. Lissie's 2010 Catching a Tiger has one track with a +3.19 dBTP and ~118k of clipping samples total. Even Eisley's Room Noises has one track with a +2.99 dBTP, with ~29k clipped samples total - there's actually more of those on Combinations at ~250k, despite a lower +2.22 dBTP. Plus, FKA twigs' LP1 has two tracks peaking slightly over 1.5, at 3.72 and 3.62 dBTP, though actual clipping samples are only ~23k. Yeah Yeah Yeahs' 2009 It's Blitz! has one track with a 3.94 dBTP, although the number of actually clipping samples is much greater on the preceding 2003 Fever To Tell at roughly 1.5 million. Which, mind you, Ladyhawke almost doubles, while Art Angels only makes to to about 1 million.)

Also, looks like there's a new record holder for my collection:

Kali Mutsa - Souvenance (2014)
Note the 6.9 million clipping samples, too...
It's an, err, interesting if decidedly eccentric album.

 

[solderdude]

I wonder what's more sound quality degrading with overs... The actual clipping of peaks or the effects of the used compressors/limiters in those recordings to get low DR numbers.
How much does added headroom to allow for ISO actually do for more enjoyment ?

 

[AnalogSteph]

Now that there is a very good question.

What I do know is:

1. It is only when brickwall limiters were fed with previously compressed + limited material that the proverbial fecal matter really tended to hit the rotating impeller. (As an aside, compressing a finished mix tends to create a terrible mush due to IMD. One of the key advantages of digital recording used to be that it wouldn't do that, unlike tape with its much higher distortion.)
2. Nonlinear operations in the digital domain - including limiters, compressors or even tape simulators - are not inherently band-limited. The harmonics and other IM products generated by multiplication stick out into adjacent repeating spectra, and since any of them is as good as any of the others, that creates aliasing. As you can probably imagine, a brickwall limiter creates plenty of those, and the more the harder it's being hit. It should thus come as a surprise to nobody that the oversampling variety was such a major advancement.

 

[dasdoing]

digital clipping is called hard clipping and can be transparent...up to a point obviously.
traditional analog clipping is now called soft clipping and will always color the sound.

 

[solderdude]

analog clipping can be just as 'hard' as digital clipping. It can even be worse when this causes a feedback loop to saturate and cause a 'memory' effect prolonging the clipping.

 

[Smaestro]

True peak is dependent on what digital signal is converted (real or simulated) into analog. To get the trueest measurement of the true peaks in a digital file you need a low pass filter that is completely straight up to the Nyquist frequency, and then straight down.

This paper is probably very interesting, as they analyse the lack of strict standards for TP measurements, the inaccuracy of 4x oversampling filters for TP measurements, and propose better filters.

E-library page - AES

AES E-Library ← Back to search

aes2.org

They also propose a more accurate filter that doesn’t have a straight frequency response. Seems interesting but I haven’t read it yet. It could have something to do with the other broadcast standards tackled in the paper.

On this page you can see their proposed filters:

iZotope, Inc - AES Paper Supplementary Materials