0dB audio

Peaks at or near 0dBFS can produce signal peaks greater than 0dB, leading to clipping on some DACs. The claim has been made that these signal peaks can be as high as +6 to +8dB.

Most recommendation aimed at avoiding playback clipping suggest limiting maximum sample values to -0.3 to -0.1 dB, but I have seen -6dB advocated.

Those more common values are probably based on something. Either that small reduction below 0dB computes to enough to avoid all clipping, the probability of clipping at those values is so small as to be negligible, or the recommendations are misguided. Who knows?

Those more common values are probably based on something. Either that small reduction below 0dB computes to enough to avoid all clipping, the probability of clipping at those values is so small as to be negligible, or the recommendations are misguided. Who knows?

The -0.1 to -0.3 values are there because some CD players have been known to make a huge fuss about the decoded value for 0dB, although to be fair, I think that this was the early ones in the main. But since you can't tell what a disc will be played on, and this difference in level is held to be inaudible, it hardly matters. I don't know where the -6dB recommendation came from - I've never seen it, but I can hazard a guess, based on what happens in the following stages:

Under very adverse circumstances, as you noted it's possible to get a much higher value than 0dB out of a D-A converter, although in general, you'd be unlikely to get more than +2-3dB. And as this is not a result of incorrect codes, the only limitation on what it does to the output is based on whether there is enough headroom in the reference voltage to be able to reproduce these levels. If there is, then through they go to the next stage. If there isn't, and the next stage has the same headroom limitation, then you will get bad clipping anywhere over 0dB anyway.

And cutting your peak level back 0.3dB won't prevent these overs. In order to absolutely guarantee to do that, you'd need about 6dB of headroom on the CD itself to prevent the output from ever rising to these levels. In all honesty I can't say that I've ever noticed or recieved notice of a problem with this myself, so I tend to stick to about -0.2dB. The +8dB figure was something of a contrivance partly based on a subsample rate trick, and I can't see this ever turning up for real.

The only material I've run into that has that potential is a tambourine I recorded this summer.  It clipped a bit when recording (live show) and attempts to use clip restoration failed because the clips were, for the most part, intersample.  I did not have multiple samples at 0dB.  It's not bad, but I can hear it.  The artist thinks it's her hand against the instrument - I don't think it is.  I can her that too and it's different.

Something that just occurred to me ... I wonder what I'd get if I upsampled the original  file to 88 or 166 KHz.  I'll have to try that later.

I think Michael Gerzon calculated that it was theoretically possible for an intersample peak to be as high as +15dBFS with certain test signals. Signals like this would never be encountered in the real world though.

Cheers

James.

I didn't calculate it, I just did an experiment with sample manipulation. I think that I used about 23 samples, and couldn't find any arrangement that would produce any significant increase over about 8 point something dB. What on earth you'd have to do to get to +15dB I really don't know - adding more samples into the pattern made less and less difference, so I stopped. What you need to do is produce the largest rate of change difference you can, and you do this by having a single trio of samples at max and min at a sampling subharmonic, and then adjusting the ones either side of them. I don't see how you could arrange for anything higher than this as an extraneous peak, but I'd be quite happy to see how it could be done - if indeed it actually can be.

Something to play with:

http://www.cheap-tracks.com/mp3/intersample_peaks.wav

30 seconds at 44.1 KHz 32 bit.  Seems to have intersample peaks on the order of 3.6dB and no (that I've found) contiguous 0dB samples.

I messed with upsampling, and was slightly surprised by the results.  It did not clip the signal - it approximated the waveform above 0dB - and with some quick listens it seems to clean up the clipped sound better than limiting and such that I did at 44.1 KHz.

I'll need to spend a little more quality time with this.

I read that article by Nika Aldrich (i think that's the name) about above 0 dbfs DA peak distortions.  

Personally, I think in most cases the clipping is for so brief of a time that it is usually perceptually forward masked (see ATRAC for info about forward masking).  

I think other times it is not perceptually forward masked, but would only be audible as a minute change in timbre if it is audible at all.  I am highly doubtful that this kind of clipping is audible for most material.  

For certain genres of music which are already loud, aggressive, and somewhat distorted, this type of peak would most definately be masked by the other sounds.  

Last but not least, the loudspeakers themselves in many cases would smooth out some of these distortions or would not be able to produce the peaks anyway whether correct or not.  Remember that loudspeaker distortion specifications are much higher than other parts of the sound chain.  Also many loudspeakers can only reproduce up to about 18-19 kHz, so many transients would be smoothed out.

Going back and looking at my sources, it would appear that the 15dB figure came from Nika and may have been created by applying an illegal stimulus (that would never happen in a properly designed system) that coincided with a particularly sensitive frequency in the DAC filter. I can find no reliable source for anything higher than a 6dB intersample peak with legal signals.

Cheers

James.