[MacacoDoSom]

  Inter-sample clipping is a HUGE problem.
 
Apple gives away software that lets you demonstrate this for yourself. It's called AU Lab and includes an AU AAC RoundTrip plugin you can use with any DAW that can use AU plugins and it is all free.
 
https://www.apple.com/itunes/mastered-for-itunes/
 
This is a fantastic tool and I urge anyone with an interest in this topic to download it and investigate for themselves.
 
Here's a video of the tool in action. The recording engineer for this particular recording was pretty shocked that I, as a consumer, would send him a scathing email regarding his handling of one of my favorite artists.
 
https://www.youtube.com/watch?v=6tb24lt09vU

It would be nice if I had a Mac...do they give away a Mac?

 

[bigshot]

A friend of mine just got an ancient G5 cheese-grater Mac with a 24 inch Apple Cinema Display on eBay for under $200. That would make a good music server.

 

[dprimary]

   
 
It is possible to detect inter-sample peaks by upsampling the track to a high sample rate (like 192 kHz), and then analyzing it. Obviously, clipping from the conversion itself should be avoided, either by using floating point samples and software that does not clip them, or by attenuating the track first.

Can you explain how that would show the inter-sample peaks. In my mind that would just add a hole lot of 0's

 

[dprimary]

  in the digital domain the samples cannot exceed 0dB full scale!!!! there is no space for it... only in 32 bit floating point...

You can't go above all 1's any ability to go over is made possible by bit shifting it down a bit or two for processing and then back up, which is part of the reason DSP is done at 24bit double precision, 32 bit floating which each has trade offs and now 64 bit floating is getting pretty common in the last few years. 

 

[RRod]

  Can you explain how that would show the inter-sample peaks. In my mind that would just add a hole lot of 0's

 
You're simulating what the output stage of the DAC would do. The 0s get added in, but then you low-pass the result just as the anti-imaging filter would do. Once you do that, you'll see the inter-sample peaks crop up. Square waves are good for this.

 

[jcx]

might want more than just 2 dB headroom: http://www.audiomisc.co.uk/HFN/OverTheTop/OTT.html

 

[limpidglitch]

What Rob and stv said, just with pictures.
A 55Hz square wave, originally at 44.1, upsampled to 88.2 and 176.4kHz, respectively.


 

 

 

[dprimary]

Ok i get it now. However with it up sampled is it not just plain digital clipping? The original sample rate would be inter-sample clipping in the analog domain with it up sampled it is now clipping in the digital domain. Which is why the codec's have trouble with the inter-sample clipping. Does anyone know if the lossless codec's have the same problem? I would not expect them to since they are doing a much simpler data compression.

 

[MacacoDoSom]

  might want more than just 2 dB headroom: http://www.audiomisc.co.uk/HFN/OverTheTop/OTT.html

great article thanks!

 

[stv014]

  Ok i get it now. However with it up sampled is it not just plain digital clipping?

 
The upsampling is only used to find out what the inter-sample peaks would be (approximately) at the original sample rate. The exact level of the peaks depends on the reconstruction filter used, so it is actually not the same for all DACs.

 

[stv014]

  might want more than just 2 dB headroom: http://www.audiomisc.co.uk/HFN/OverTheTop/OTT.html

 
With synthetic test signals, the inter-sample peaks can be made very high, but such peaks are unlikely to occur in real music that is not already heavily compressed and clipped (and in that case, it is not even obvious whether having the inter-sample peaks above 0 dBFS is correct or not, as the peak limiting/clipping algorithm running in continuous time would not have produced peaks that exceed 0 dBFS).
 
It can be a real problem with some special signals used in measurements, for example, a maximum length sequence can give peaks above +8 dBFS even with a filter that simulates a response that can be found in real DACs. And a tone at exactly Fs/2 that is inverted after some point can reach even higher levels, depending on how close the filter is to the ideal response (infinitely fast roll-off).
 

This sample that originally had a peak level of only -20 dBFS at 44100 Hz sample rate would be clipped at 192000 Hz in an integer format (although it was upsampled with a FIR filter that is millions of samples long; in theory, the peak could be made infinitely high, but its level is a logarithmic function of the IR length of the filter).

 

[stv014]

  What Rob and stv said, just with pictures.
A 55Hz square wave, originally at 44.1, upsampled to 88.2 and 176.4kHz, respectively.


 

 

 
Another solution is to delay the input by half sample with sinc interpolation. This is basically the same as upsampling to 88.2 kHz, except the output does not include samples that would be the same (with linear phase filtering) as the input signal. Inter-sample peaks are not necessarily exactly halfway between the samples, as it can be seen above, but the result is reasonably accurate.

 

[dprimary]

   
The upsampling is only used to find out what the inter-sample peaks would be (approximately) at the original sample rate. The exact level of the peaks depends on the reconstruction filter used, so it is actually not the same for all DACs.

If for some reason you do have to up-sample, for example bringing music with heavy inter-sample clipping in into a film and for some reason the film is done at 96k or 192 would it now not be clipping in the digital realm? Hopefully someone would notice it is clipping during the conversion but I would not count on it. 

 

[MacacoDoSom]

 

   
The upsampling is only used to find out what the inter-sample peaks would be (approximately) at the original sample rate. The exact level of the peaks depends on the reconstruction filter used, so it is actually not the same for all DACs.

If for some reason you do have to up-sample, for example bringing music with heavy inter-sample clipping in into a film and for some reason the film is done at 96k or 192 would it now not be clipping in the digital realm? Hopefully someone would notice it is clipping during the conversion but I would not count on it. 

some SRCs have an option to prevent clipping...

 

[arnyk]

  Most CDs have 0 Db FS witch most of the time results in having intersample peaks (TP) of +2/+3 dB, would this be an issue with most modern DACs? or they always have enough headroom to to fit these? and about the amplifier?
any thoughts?

 
Just because there is clipping does not mean that there is necessarily an intersample peak. In fact most clippied  fiiles are free   of the them or only have a few.
 
Most music can be clipped for brief periods without it being audible. Temporal and spectral masking is to blame.
 
Creating an ISP takes a certain configuration of data that usually includes for example a sharp rise followed immediately by a sharp drop. Most clipping fails to do this because the cliipped period is too long.
 
That does not meant that ISPs don't ever happen. I've seen a goodly number of them where all samples involved were well below clipping.