Re: onset detection

[Andrew McGuiness <andymcguiness@xxxxxxxxxxxx>]

Thanks, Bob. I had another look at your ""Smooth" Rhythms" paper and found it quite interesting. 
In fact, the negative-going zero crossing of the 2nd derivative of energy is just the highest
point of the peak of the 1st derivative (ie. the 'steepest slope' that I mentioned in the second
par of my post: "Having found the general area of the onset, the script looks for the
 steepest point in the energy curve of the upper partials of the sound, to identify the precise
point of onset." Your method will work fine on smooth energy curves with one peak per onset, but
not for fine discrimination on an audio file which is a bit messy (say, one instrument in a
drumkit, recorded in an ensemble situation). I was pleased to find that you located the P-center
at the peak of the 1st derivative though.


> 
> Andy,
> 
> 	For a counter example, you might look at an old paper of mine, =20=
> 
> =93Smooth=94 Rhythms as Probes of Entrainment,=94 _Music Perception_ 10 =20=
> 
> (1993): 503-508 (<http://faculty-web.at.northwestern.edu/music/=20
> gjerdingen/Papers/PubPapers/SmoothRhythms.pdf>). There the perceived =20
> onsets occur at positive-to-negative zero crossings of the second =20
> derivative of energy. This would be in a region of all positive =20
> slope, if I'm not misunderstanding your method.
> 
> Best wishes for the Holidays,
> Bob Gjerdingen
> Northwestern Univ. School of Music=
> 

hi Nick - Of course I had already looked (although briefly) at the Mirex stuff, but it didn't meet
my needs. I needed a finer-grained discrimination than 50 milliseconds. On the other hand, I don't
need to find onsets in real time (as you did with your work) and I can put up with a few false
positives (which I can remove by hand). You're prob right, though, about there not being a simple
relation between onset and PAT - I just thought there might have been something I hadn't come
across yet.

Andy
> ------------------------------
> 
> Date:    Wed, 26 Dec 2007 19:06:50 +0000
> From:    Nick Collins <nc272@xxxxxxxxx>
> Subject: Re: onset detection
> 
> Hi Andy,
> 
> You'll probably want to look into the large literature on onset detection 
> algorithms, perhaps starting from the MIREX competitions?
> 
> I've linked the most recent here:
> 
> http://www.music-ir.org/mirex/2007/index.php/Audio_Onset_Detection_Results
> 
> Onset detection based on overall energy change over 100 msec blocks is not 
> necessarily a very general strategy. Further, p-centre is potentially a 
> rather complicated phenomena with a further literature and not to be 
> confounded with physical onset detection. Not sure you'll find a simple 
> relation between perceptual and physical onset based on a simple single 
> band  energy change detection function, especially over a variety of 
> timbres.
> 
> best,
> Nick
> 
> 
> 
> --On 26 December 2007 11:31:05 -0600 robert gjerdingen 
> <r-gjerdingen@xxxxxxxxxxxxxxxx> wrote:
> 
> > On Dec 25, 2007, at 5:57 AM, Andrew McGuiness wrote:
> >
> >> First, the script looks for negative slope in the energy curve of
> >> the sound, with energy calculate by averaging 100 millisecond
> >> blocks, to eliminate very local fluctuations. If the sound file is
> >> recorded clean, with no spill, this works well to identify the
> >> general area of the onset.
> >>
> >
> 
> ------------------------------
> 
> End of AUDITORY Digest - 25 Dec 2007 to 26 Dec 2007 (#2007-295)
> ***************************************************************
> 



      Make the switch to the world's best email. Get the new Yahoo!7 Mail now. www.yahoo7.com.au/worldsbestemail