Re: AMDF shortcuts for SOLA time stretching?

[Mattijs Kneppers]

Hi,

I'm new on this list, and happy to read the interesting discussions that are going on here (although I'll have to acquaint myself with some of the jargon).

I just finished a time stretching implementation, more or less PSOLA, that you might be interested in. I think the quality comes close to the zplane stuff.

You can find it here, it is done in MaxMSP 5 and it's open source: http://www.cycling74.com/twiki/bin/view/Share/MattijsKneppers

Best,
Mattijs

----- Original Message ----

> From: James Chandler Jr <jchandjr@...>
> To: A discussion list for music-related DSP <music-dsp@...>
> Sent: Wednesday, April 23, 2008 7:26:37 AM
> Subject: Re: [music-dsp] AMDF shortcuts for SOLA time stretching?
>
> Hi Ross
>
> I look at all local maxima, both positive and negative. Autocorr of all peaks
> against all other peaks (within the max/min autocorr time limits) seems to give
> 'the very worst matches' along with 'the very best matches'. A very bad score
> might happen when doing correlation between a positive peak and a negative peak
> about 1.5 wave periods distant, or whatever.
>
> Quality-scaling of the array of AMDF values is a simple formula which
> incorporates the very worst score found, so it is nice to find the really bad
> ones along with the good scores.
>
> //high-level check after calculating an array of AMDF scores in a time slice
> BestBeatPeriodicity := (AutoCorrMaxVal - AutoCorrMinVal) / AutoCorrMaxVal;
> if BestBeatPeriodicity < 0.7 then //don't mess with bad periodicity loops
>   goto Skip1; //nothing to see here folks, move along
>
> Using peaks 'seemed like a good idea at the time'. I think the reasons not to
> use zero-crossings, were mostly:
>
> 1. If a file contains low level noise, perhaps the low-level noise would
> modulate the locations of zero crossings more than it would modulate the
> locations of loud peaks.
>
> Viewing wave-animations or oscilloscope traces of some instruments, you see
> higher harmonics 'rippling' or 'slip-sliding' across the tops of lower
> harmonics. This can modulate the peak locations, but I think it is far more
> likely to modulate zero crossing locations.
>
> 2. More importantly, some signals have significant long-duration DC shifts,
> usually near note-on locations. Notorious on Guitar or Electric Bass, though it
> can also be seen in Vocals or Sax, whatever.
>
> For instance, a guitar note can get DC-shifted A LOT for ahile after the
> beginning of the note. This plays hob with zero-crossing locations, but does not
>
> significantly affect the time-location of the local peaks. Even if an aggressive
>
> note-on 'bottom peak' happened to get shifted above the zero line, you still can
>
> detect the time location of a severely-shifted 'bottom peak' by marching thru
> the samples and marking all slope reversals.
>
> Long ago I tried getting rid of transient DC shifts with sharp hipass filters,
> to make zero-crossing detection better. But that didn't help much if at all. If
> anything, it caused note-ons to have a 'double bounce' in DC level, making it
> worse.
>
> ===
>
> The nonrealtime beat detection provides some luxuries that would be difficult in
>
> a realtime plugin that never knows what is coming next.
>
> This code has an analysis option for 'fine-tuning' the short-loop detection,
> once the best 'gross' locations are found. The fine-tune option is used for only
>
> some instrument types. Given the best-scoring 'gross' short-loops, it will do
> AMDF sample-by-sample over a time range of +/- 1 semitone, to find if there is a
>
> better score in the vicinity of that 'best score'.
>
> I expected this fine-tune option to be most useful with instruments using pretty
>
> short autocorr ranges, such as Sax. Strangely enough, the fine-tune short-loop
> discovery was fairly useless for short-autocorr range mono instruments, but a
> noticeable improvement for long-autocorr range chord instruments such as
> acoustic guitar. Not expected, since a short-loop on a chord is several
> milliseconds, and one might not expect a change of a couple of samples' distance
>
> to matter much.
>
> Debug-watching the fine-tune loop, it looked like perhaps 20 percent of short
> loops would get modified from the 'gross detection length', and typically only
> by one or two samples. It seems surprising that the 'gross peaks' distance is so
>
> frequently near optimum.
>
> ===
>
> In previous programs, for time-domain pitch detection or stretching, I would
> start at some anchor point, and do a loop of autocorrelation or AMDF at semitone
>
> distances (to save compute time). Perhaps 24 autocorrelations for a distance of
> 2 octaves. Then I would do a second fine-tune loop for +/- 0.5 semitone,
> sample-by-sample, centered on the best gross score.
>
> That didn't work any better than the current method of just comparing all peaks
> to all other peaks, and the peak cross-correlation perhaps runs faster.
>
> ===
>
> If a note lasts a long time, maybe one second or more, the current
> CompareAllPeaksAgainstAllOtherPeaks, requires a potentially big temporary score
> array.
>
> For instance, if you wanted to accomodate an array of a max of 1001 peak
> locations (lots of peaks in a long note), then a fully-filled out array of
> cross-correlations, I think, would be a triangular number-- (n^2 + n) / 2. So
> for 1001 peaks, a fully-populated cross-correlation array would need 501501
> elements?
>
> But since there are max and min autocorr distance limits, and the
> cross-correlation array doesn't have to be fully populated, it isn't quite that
> bad.
>
> Set some debug values to fire if I ran out of array space, and in practical
> usage, the array of peak locations had to be inched up to 8000 to 'never fail'
> on very long notes. But a cross-correlation array size of 500,000 elements
> turned out big enough to never fail in practice. Perhaps it could be
> significantly smaller than that, but I got tired of fiddling with it.
>
> jcjr
>
> ----- Original Message -----
> From: "Ross Bencina"
> To: "A discussion list for music-related DSP"
> Sent: Tuesday, April 22, 2008 6:41 AM
> Subject: Re: [music-dsp] AMDF shortcuts for SOLA time stretching?
>
>
> > Thanks for your usual generosity in sharing all the interesting details James,
>
> > I didn't find that tedious at all.
> >
> > I found the "Good Short Loops" bit especially interesting.. I'm still not sure
>
> > how you find it easier to find peaks instead of zero crossings for
> > autocorrelation anchor points. When you say 'peaks' do you mean all local
> > maxima? or maximum point between adjacent zero crossings? Is your resoning
> > that the peaks provide a larger set of potential anchor points to work with?
> >
> > Best wishes
> >
> > Ross.
>
> --
> dupswapdrop -- the music-dsp mailing list and website:
> subscription info, FAQ, source code archive, list archive, book reviews, dsp
> links
> http://music.columbia.edu/cmc/music-dsp 
> http://music.columbia.edu/mailman/listinfo/music-dsp

--
dupswapdrop -- the music-dsp mailing list and website:
subscription info, FAQ, source code archive, list archive, book reviews, dsp links
http://music.columbia.edu/cmc/music-dsp 
http://music.columbia.edu/mailman/listinfo/music-dsp