[SP Wild]

Right on gregorio. Samples within music of the past decade are of much higher quality than before, despite the loudness wars. Audiophiles with their golden ears wouldn't be able to figure that out.

 

[Roseval]

   
 These inter-sample peaks are commonly no more than about 0.3dB higher than the 0dBFS limit.

Might be a bit to low.
Benchmark measured even +3.5 !
https://benchmarkmedia.com/blogs/application_notes/13545433-audio-that-goes-to-11

 

[gregorio]

  Might be a bit to low.
Benchmark measured even +3.5 !
https://benchmarkmedia.com/blogs/application_notes/13545433-audio-that-goes-to-11

 
There is no fixed value for IS peaks and no way of predicting when they will occur or how large they will be just from listening to the material. With experience one can guess, from listening, the passages where it's most likely to occur but not how much and not with a particularly high success rate. IS peaks can be as high as +30dB or so in theory, although you can only get them that high deliberately, using Dirac impulses.
 
The delivery requirements of much of the work I do specifies peak level in dBTP, so for several years I've seen and measured first hand the occurrence of IS peaks on a variety of audio content. In my experience, the majority of the time, ISPs are no more than +0.3dBFS, ISPs greater than +1dBFS are relatively rare and only a couple of times have I seen an ISP hit +3dBFS. I suppose if one specialised in certain types of audio or certain musical genres, recorded or produced in a certain way, then seeing ISPs that high would be more common but IME, IS peaks that high are very much the exception rather than the rule. It's worth noting, again IME, that it's also rare to come across content where there are no ISPs (IE. Where the max dBTP level does not exceed the peak dBFS level).
 
Some caveats:
1. The lower the sample rate, the greater the magnitude the ISPs are likely to be. IE., ISPs are likely to be slightly lower with a sample rate of 48kHz than with 44.1 and potentially lower again at 96kHz.
2. Likewise, the higher the sample rate of what we're using to measure the ISPs (IE. A true peak meter), the higher the magnitude the ISPs could measure.
3. There is no standardised method of measuring IS peaks (dBTP). ITU BS 1770-3 requires a compliant TP meter to employ a minimum of 4x oversampling. I personally use a dBTP meter which employs 9x oversampling.
4. Some processes may employ far higher oversampling rates and may therefore encounter higher ISP magnitudes. Which is why the EBU R128 specifies a max level of -1dBTP and the ATSC A/85 is even more conservative with a -2dBTP max (both based on the ITU's 4x oversampled requirement).
5. Although there may be some, I'm personally not aware of published data comparing dBFS and dBTP values for a variety of content, so I have no evidence to substantiate my claims, only personal experience (and a few conversations with some other audio engineers on the subject).
 
G

 

[castleofargh]

looking for peak meters with different upsampling logically change the readings. I wish all peak meter plug ins would clearly state the value used.
I noticed that intersample clipping was rarely a problem for with CDs. it happens of course, but I usually had less than 1db and can't say that it sounded obvious by ear. but for mp3, long before I had any idea what intersample clipping was, I've spent the longest time not understanding why a track could sound bad right after I ripped it(some percussion would feel distorted or whatever), but after applying replaygain most problematic songs were good again. I ended up with foobar's preamp at -3db as some general patch for that unknown problem.
 
and I always wonder how many people spit on mp3 saying it sounds like crap for that reason?

 

[gregorio]

  [1] looking for peak meters with different upsampling logically change the readings. I wish all peak meter plug ins would clearly state the value used.
[2] and I always wonder how many people spit on mp3 saying it sounds like crap for that reason?
[3] I ended up with foobar's preamp at -3db as some general patch for that unknown problem.

 
1. To be honest, it's not really of any importance to anyone, except in discussions of this type! When I was looking for a TP meter I was only concerned whether or not it was ITU 1770 compliant (4x oversampled), it wasn't until sometime after I'd already purchased one that I discovered it actually oversampled by 9x. The fact that it is 9x oversampled is only slightly useful, in that it may read some ISPs as marginally higher (than a 4x oversampling meter) and therefore ensure that I'm always within specs, assuming of course that a particular QC ingest department are not using a TP meter with an even higher sampling rate! But even in that case, I can still legitimately argue I'm within EBU or ATSC specs.
 
2. With MP3 there's the added problem of different algorithms to create valid MP3 files. Certainly some of those algos don't provide headroom and effectively just clip the ISPs, which as you say has no doubt affected some tests/quality judgements. Another area covered by my #4 caveat above is sample rate conversion, which commonly requires very high oversampling (to an integer ratio of both the original and converted rates). Most SRC processes create headroom to avoid clipping ISPs and the result is usually a converted file which is a fraction of a dB lower than the original. A fraction of a dB is inconsequential in just about any practical application but maybe enough to influence the outcome of an AB test between say a hires file and that same file downsampled to 44.1.
 
3. There's no easy/perfect solution to this. To be absolutely certain of never incurring ISP clipping, you would need to set your preamp to -30dB but of course that's going to screw-up the gain staging! Getting a TP meter and then setting your preamp to -2dBFS below that TP reading will have you covered in pretty much any conceivable situation but of course measuring each track and adjusting the preamp level accordingly is tedious and impractical with a large number of tracks. So, I would say that your -3dB setting is probably a good balanced choice, which will eliminate the chance of ISPs clipping in the vast majority (though probably/possibly not all) of cases.
 
G