17

I am relatively new to the bird bioacoustics world (I'm a primatologist by trade but starting a side project with bird vocals from passive acoustic data I've got from Madagascar) and was a bit confused at the presence of both BirdNET and Merlin in the space. They're both out of the Cornell Lab of Ornithology (Lab of O) and both doing automated detections of bird calls, but what have peoples' experiences been using one or the other? Is one "better" or more appropriate than the other for different use cases? (I realize "better" is obviously very subjective)

If I'm understanding what I've found online already, there are different algorithms under the hood and different training datasets? Also is there a directory of supported species for either? Seems like Merlin is mostly for Global North species but I can't find a list anywhere. If that's the case, obviously Merlin wouldn't do me any good with Malagasy species. I emailed the Lab of O asking about this but haven't heard back so figured I'd tap into this hive mind to see what other folks' experiences have been.

[Edit] It was also brought up that it would be good to know if the algorithm on both these apps is embedded or performed on the cloud and if it is open-source or not.

Thanks for any insights/clarifications on comparing the two!

4
  • 1
    Hi @Carly Batist, could you please extend Cornell Lab of O please?
    – Thejasvi
    Jul 20 at 14:19
  • 2
    It would also be interesting to know for both these apps if the algorithm is embedded or performed on the cloud and if it is open-source or not.
    – Maxime BRU
    Jul 20 at 14:25
  • 1
    Hi @Thejasvi sorry about that! Just edited it Jul 20 at 16:21
  • @MaximeBRU I agree I'd like to know those answers as well! I'll edit the post to include those points Jul 20 at 16:22

1 Answer 1

15

Based on BirdNET's recent update, my estimation is that this space is changing rapidly, so take my limited understanding with a grain of salt.

Personally, I prefer the Merlin app for exploratory purposes like teaching myself species ID from my phone in the field in North America. However, when it comes to large-scale audio processing of terabytes of audio on my computer, BirdNET is the best off-the-shelf tool I'm aware of. I think it's important to distinguish between "BirdNET the phone app" and "BirdNET, the platform for scientific audio data processing" (see BirdNET-Analyzer github), the latter of which you can install on your computer and use for scientific research.

The Merlin app does real-time ID -- you don't have to select the portion of the spectrogram you want it to identify (which is currently the case for BirdNET). Merlin doesn't require cellular data once you've downloaded it (BirdNET does; it submits the selected portion of your recording to their servers for ID). I think it's much easier to access my audio recordings from Merlin (whereas accessing recordings from the BirdNET phone app is still more of a process). For Merlin, I have found this eBird help center article useful. That link also has a "Media Best Practices" section that explains how users can upload their Merlin recordings to eBird (I don't believe it is currently possible to upload them directly from the app, but I could be wrong).

This Golden Gate Audubon Society article from July 2021 reviews a few of the differences between the Merlin and BirdNET phone apps, but some of this is already out of date given that BirdNET recently went through a large update that expanded it's ID capacity from 984 North American species to ~3000 species worldwide.

From the article:

Holger Klinck, one of the researchers working on BirdNET, says, “Fundamentally, both apps are using similar deep neural network approaches. One difference is that the Merlin’s [Sound ID] algorithm is exclusively trained with data hosted by the Macaulay Library. BirdNET uses additional (non-Cornell) data sources.”

BirdNET's algorithm is described in Kahl et al. 2021, whereas I'm not able to easily locate information on Merlin's underlying process (Edited to Add: there are some interesting insights from the Merlin developers buried in Twitter threads, for example here, here, here... perhaps a paper describing Merlin is on the way?). My understanding of the BirdNET phone app, based on the app tutorial section, is that each selected portion of a recording gets sent to the BirdNET servers in Chemnitz, Germany in order to analyze the recordings and identify species (there is an option to "claim authorship" for your recordings). The BirdNET App also has a "submit feedback" button that allows a user to verify whether a species identification is correct or not, which presumably could also be used to update their models in the future.

A recent journal article from some of the BirdNET team (Wood et al. 2022) implies forthcoming updates to future versions of the BirdNET app, which hopefully will enable it to be a more user-friendly community science tool:

"A forthcoming app user data analysis portal will allow users to analyze their own observations in greater detail than the current species list and “explore your area” features, an improvement that will enable users to pursue their own research questions... An optional “point count mode” with which users could submit 3- to 5-minute continuous soundscape recordings would allow communities to document an estimated 5% to 15% of the local bird species per week even with minimal participation 10. Moreover, these acoustic point counts would yield analytically valuable nondetections of the remaining species.

And later on:

"Raw observations are not yet publicly available due to the substantial challenges to hosting tens of millions of observations and audio files; sharing observations only (i.e., prediction scores and metadata but no audio) is simpler but prevents validation of observations and thus leaves the data vulnerable to misinterpretation. Detailed data usage guidelines that will facilitate widespread open-access data sharing are forthcoming."

EDITED TO ADD: more differences between Merlin SoundID and BirdNET are discussed here in this talk by one of the SoundID developers at 57:24: https://www.youtube.com/watch?v=mVhe-MiVwLE&t=3444s

Re your question: Is there a directory of supported species for either?

3
  • 2
    That JSON file mentioned above is NOT the list of supported species, it is all birds supported by eBird, which BirdNet does not support all of now. This is the current labels list: github.com/kahst/BirdNET-Analyzer/blob/main/checkpoints/V2.2/…. I've added that link to the original comment! Aug 30 at 22:03
  • 3
    Thanks for catching that, @CarlyBatist! Also, BirdNET was updated to V2.2 about two weeks ago. In theory this comment would need to be updated periodically to reflect any time the model version is updated; hopefully readers can stay aware that this is an actively developing space and links may not be stable for long!
    – Cathleen B
    Aug 31 at 17:48
  • 2
    Windows users that find the BirdNET-Analyzer installation instructions at github.com/kahst/BirdNET-Analyzer intimidating can find simplified installation instructions at ravensoundsoftware.com/knowledge-base/birdnet-analyzer. Sep 1 at 4:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.