[aubio-user] Offset detection

[Lukasz Tracewski]

> I assume you mean free and open source software, not freeware?
> 
You are right of course; somehow 'freeware' got associated in my mind with open software, though naturally they have little in common. I only just begun my adventure with open source; here is my project's page: https://github.com/tracek/Ornitokrites It's still very immature and under heavy development. Most important part is still missing: machine learning.

Your library is one of its crucial ingredients. If things work as intended the program will be used for aiding efforts in nature conservation (initially kiwi protection). Of course it will properly cite and credit all used publications and libraries. Please let me know if there are any licensing issues.



> It would help to have an example of the kind of annotations you want.
> 
In essence I need to know beginning and end of anything that could be a bird call - if this is what you are asking. Examples will follow.



> You could just use aubiopitch. by setting the confidence and silence
> thresholds appropriately, you should get 0.00 when the bird is not singing.
> 
Thanks! I will try that.




> By the way, I would be very interested to have a small sample of songs
> and the 'ground-truth' annotations you are looking for.
> 
Here you will find some examples:
https://drive.google.com/folderview?id=0B46Cy3WOKgyJb216cHYwSE1yeTA&usp=sharing

Three folders contain respectively:
- Original audio recording
... and output of my script:
- Spectrogram with beginnings and ends of a probable audio features (e.g. a bird call) marked by vertical lines. Currently the latter are created in a naive fashion: [detected_onset - 0.2 second , detected_onset + 0.8 second], giving 1 second-long samples. Of course these hard-coded intervals won't do in general.
- Shorter recording containing only audio features. This one is created by extracting areas marked on a spectrogram and performing noise reduction: high-pass filter and spectral subtraction. The latter creates so-called musical tones (tin-like sounds) and one of my goals is to make it smarter (like "noise removal" in Audacity).

For convenient download I also included a zip file with all three folders. 

My aim is to identify which recordings contain kiwi calls. By the way, if you are interested in how kiwi sounds:
https://www.kiwisforkiwi.org/about-kiwi/kiwi-facts-characteristics/kiwi-calls/
The site also contains a lot of other great kiwi material :-).


Regards,
Lucas


> 
> 
> Regards, Paul
> 
> >
> >
> >Cheers, Lucas
> >
> >On 04/01/14, *Paul Brossier * <piem at piem.org> wrote:
> >>On 04/01/2014 06:00, Lukasz Tracewski wrote:
> >>>Hi,
> >>>
> >>>Aubio has a plethora of great onset detection methods. How about
> >>>offset detection? I can imagine that running algorithm on
> >>>reverse-ordered data should provide me with offsets, but maybe
> >>>there is a smarter approach that I am missing.
> >>
> >>Hi Lukasz,
> >>
> >>There is no offset detection method in aubio (yet). The onset
> >>detection functions will sometimes peak during the transient sound
> >>created by the release of a note, but these peaks are expected be
> >>small compared to the ones produced by note onsets.
> >>
> >>So far, offsets times are assumed to be where the level of the
> >>signal drops under a given threshold. aubionotes for instance uses
> >>a simple silence gate to determine the end of notes.
> >>
> >>What kind of offsets are you looking at? Do you have a specific
> >>algorithm in mind?
> >>
> >>There is the work of Emmanouil Benetos and Simon Dixon, who use a
> >>Hidden Markov Model to determine whether a pitch track is active
> >>or not:
> >>
> >>http://www.eecs.qmul.ac.uk/~simond/pub/2011/Benetos-Dixon-ICASSP2011.pdf
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> Best, Paul
> >>
> >>
> >>>
> >>>Thanks, Lucas
> >>>
> >>>
> >>>_______________________________________________ aubio-user
> >>>mailing list aubio-user at aubio.org
> >>>https://lists.aubio.org/listinfo/aubio-user
> >>>
> 
> 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.aubio.org/pipermail/aubio-user/attachments/20140110/23bf6032/attachment.html>