[kiproping]

I wanted to do something similar to this, then I started doing some research on birds in general, and those in my locality, then I started learning about Audio and spectograms and Nyquist Theorem and many other interesting audio stuff.

Then I started going through the Intro to Conservation Bioacoustics by Cornell course, and started watching Bioacoustic Talks by the K. Lisa Yang Center cornell center.

And now I am almost at the point where I cant start manually tagging audio sets, for target species so that I can train custom classifiers to identify birds in Rwanda which are poorly detected by birdnet.

TLDR: Being jobless can lead you into interesting ventures.

* Nyquist Theorem. https://www.youtube.com/watch?v=IZJQXlbm2dU

* Intro to Conservation Bioacoustics https://www.birds.cornell.edu/ccb/pam-materials

* Bioacoustic Talks https://www.youtube.com/@CornellSounds

[jna_sh]

Thanks for sharing these resources and your story! I followed a very similar path, and ended up doing a biodiversity related MSc, with my dissertation being a custom classifier for poorly detected species in Príncipe. BirdNET and Perch are phenomenal achievements, but struggle in regions where, ironically, most of the world’s biodiversity is. What you’re doing for Rwandan species is so important!!

[kiproping]

thanks, how is your work going. Any interesting papers you did, what is an area you think needs more research.

[jna_sh]

I have alas not published it yet, but I really should. How about you? For me, on the research front, I’m very interested in methods that can be sustainably applied in remote, resource constrained locations. Heavy cloud dependent workflows to adapt a huge foundation model just aren’t practical on an island that doesn’t have 24/7 electricity and sporadic connectivity.

[is_true]

Do you think the same could be used for other cases? I'm thinking about detecting problems with cars (vehicles) just by the noise they make

[kiproping]

I know google has general sound classifiers like Yamnet, trained on youtube data but they are not very good for specific usecases. So you would have to create a custom model for you usecase.

- https://www.tensorflow.org/hub/tutorials/yamnet

[smalldude]

How far has your thinking taken you? This has piqued my interest on how many fields audio can be used to solve problems

[is_true]

just thinking, I've talked with a mechanic but he told me that now when they connect the car to a computer they almost always find anything wrong with a car, that and the experience they have they almost always know what's wrong.

I think sound + location could be really interesting, because you can filter parts of the car that could be making noises that are similar knowing where the mic is.