Re: AUBIO for onset detection -- in FreeWheeling?

[Paul Brossier <piem@altern.org>]

Hi JP, hi all,

Thanks for your mail and sorry for the late reply, I was away for
the past 10 days.

On Wed, Feb 16, 2005 at 12:06:06PM -0700, JP Mercury wrote:
> Dear Paul,
> 
> I've been playing with your aubio labeling library- it's good! I tried the
> onset and note detection features and they got me thinking.
>
> I see a potential application for FreeWheeling that I wanted to run by you.
> I'd like to autodetect a first loop when a player starts to repeat a phrase.
> What I thought is to run an input through AUBIO and grab the onsets, then to
> compare the audio at previous onsets to find the most likely beat length-- if
> a strong correlation is found over several repetitions, a loop would be
> automatically created. This would allow an improviser to 'gel' a loop just by
> repeating it. The timing of those first loops can be tricky and I thought it
> might be a good aid. 

I also thought there is a good potential here, and these sound
like a great idea for FreeWheeling.

> I wonder how reliably such a system could work? AUBIO seems to reliably detect
> the onsets from my rhythms, with only a few stray onsets. I noticed that it is
> quite tolerant of different levels. Combined with an autocorrelation to
> determine likely loop points, it could work. Do you think it would be a good
> approach?

The loudness of the audio input should indeed not change the
output results, as the onset peak picking is adaptive and does
not take the amplitude in account. Most of the false positives
can be avoided by increasing the threshold (and 0.1 is actually a
bit low for a default value).

A beat tracking algorithm should soon be included in the library.
The process indeed uses autocorrelation of the onset detection
function to determine the most likely beat.

You can read the following paper to learn more about this.
http://www.iua.upf.es/mtg/ismir2004/review/abstract.php?idPaper=226

> One caveat, aubioonset and aubionotes don't seem to work with my default 64
> sample size and 2 jack buffers- if I up to 1024, OK. Is this because you are
> doing FFT work and not buffering?

Arg, this bug probably occured in version 0.1.8, where i decided
to remove a buffer copy that looked useless. I will get that
fixed soon.

> Once again, thanks for your work on this.
> 
> -Mercury

Thank You for your feedback,

Best regards, piem


-- 
To unsubscribe, send mail to aubio-unsubscribe@piem.org.