discussion / Acoustic Monitoring  / 9 June 2022

Bird Acoustic Solution 

Hello Folks, I work with the Dept of Conservation, New Zealand (NZ) government. We are eagerly looking for a Machine learning or AI solution to identify the NZ birds (and other species in future). Department has historical audio/acoustic recordings gathered by our rangers; we want a solution to scan these acoustic files to identify the birds. This solution is essential for us, and the department is willing to invest a reasonable amount in enhancing/developing the existing/premature solution, which can also be made available to other conservation groups in NZ. Government has a target of predator-free NZ by 2050, and this solution and similar others in future will help us reach that target. 

I am new to Wildlabs.net and have seen some cool and advanced technologies used in conservation space I am extremely excited to implement them into our organization. I shared some of these capabilities in our organization, and there was an interest and curiosity. In that direction, I have been asked if we could find a solution to our audio/acoustic files.  

Please let me know if you have a solution to this problem and are willing to collaborate to take the solution to the next level. My email id is [email protected]; feel free to email or message me for discussion. 

Thank you so much.

Jitendra 




Hi Jitendra. 

I recommend checking out these videos from the WildLabs TechTutor series and perhaps contacting the speakers for advice and recommendations:

Jamie Macaulay: How do I analyse large acoustic datasets using PAMGuard?

Zephyr Gold & Marconi Campos: How do I use pattern matching to label acoustic data with RFCxArbimon?

Good luck with the project. Hope to hear more about it as it progresses :)

Akiba


 


 

There are a bunch of different options for detecting calls in audio data, from proper statistical platforms such as R/Python, to bespoke software such as Arbimon, Kaleidoscope & Raven. Edge Impulse also an online ML model-building interface, but this is more focused on then deploying the models onto devices for edge computing. Arbimon has template matching features that are a good way to start finding detections to build a training dataset, I have used it for this in the past. Arbimon is online & free. Kaleidoscope has a clustering function which is again a good first step to start picking out the low-hanging fruit of detections so to speak. It's a desktop app, but this is not free ($400/yr). Raven also has some automated features -  template & band-limited entropy detectors. It's also a desktop app and not free ($100-$800 depending on 1-year or permanent license and whether non-profit or not; not sure where a government agency would fit into that). 

There is always the ubiquitous split between biologists who traditionally are taught to use R and tech/computer folks who are taught to use Python, but for ML, Python's ecosystem is really well set up. Not sure what the level of programming you/your dept has, but there are a TON of free resources online for learning it if you were interested.  

Relevant Python bioacoustics packages potentially of use - Acoustic_Indices, scikit-maad, Ketos, OpenSoundscape (as well as the obvious ML ones such as TensorFlow)

Some R packages as well -  soundecology, bioacoustics, monitoR, warbleR, gibbonR

@tessa_rhinehart has created a fabulous list of bioacoustics software that you can find here: https://github.com/rhine3/bioacoustics-software

You can also turn to articles that have already done similar things and reach out to the authors to discuss their methods. I've got a (totally un-exhaustive) list of papers on passive acoustic monitoring, with a section on 'analyses' that you might find useful to start with; I can email it to you if you'd like. Working on a PAM training materials page on my website that it will be available at shortly as well (will post the link to Wildlabs when it's live!).

Hope this is helpful!