I have used RAVEN software auto detect filters to identify (label) the notes of a single bird species that imitates other birds. Thus I'm working with a large data set (hundreds of thousands of notes within thousands of songs) and am attempting to label all the sounds (notes, syllables, phrases) accurately within the songs of
a bird species and found:
1) it is very time consuming to tweak the parameters to even get it to 'work'
2) when it is 'working' it returns with numerous false positive and negative results
I'm curious if those interested in identifying (labeling) large data sets of animal sounds have found a software that will sift through the spectrograms with accurate identification (labeling) of sounds?
Currently I'm using visual and aural inspection (human) of spectrograms to do this, which amazingly seems to be the only way to achieve accuracy in this task. It's incredibly time-consuming but I appear to be quicker at doing it 'old school' than by using automated (computer based) methods.
Cheers!
Brandi Gartland
M.S., Doctoral candidate in Animal Behavior
University of California, Davis
Hi Brandi,
I am not an expert on automatic classification of syllables and do not have experience with Raven used the way you describe. But I did some labelling on automatic recordings and I can share some thaughts.
1) Quality of recording is very important. Bad quality recordings will be always pain. So first try to figure out if your problems are related to sound recording quality.
2) If you need to classify end describe each element you could try to read this paper.
Article “Ficedula”: An open-source MATLAB toolbox for cutting, segme...
It is about you are asking for and there might be also mentions of other software tools. I know there were a few but cannot remind references.
3) Are you using the calls or syllables of other species to find imitations in your focal species? This might be tricky as imitations do not need to match perfectly forcing you to relax the searching parameters. I am not 100 percent shure but I doubt there are currently recognizers that would do something your ear is doing = this sounds kind of like an imitation of "species A". But on the other hand, you do want to avoid subjective bias... It could be possible to try to be very strict on recognition parameters. You assume that imitation is exactly the same as model. You will miss a lot of "sounds kind like a species X" but you might end up with very few but precise imiatations. A good starting point
4) Yes, sometimes it is best to do it all manually. But maybe think if you need to do all material at once or maybe start with some preliminary analysis on limited data to get precise idea on what you want to get from the data before you process all the material.
Good luck!
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Software