Settings to match Merlin Sound ID?

[farski]

My birdnet-PI and birdnet-go detections have always been significantly lower than simultaneous tests run against the Merlin iOS app. In my experience, the Merlin app has not had any issues with false positives; when it picks up a detection, I have heard the bird.

I've always worried that my microphone setup for birdnet has been the issue. I'm not super happy with the setup, but I haven't had the time to do a full rethink and get it upgraded.

But I just ran a test where I had Merlin listen to the audio stream coming from birdnet-go, and Merlin was still able to pick up way more detections than birdnet-go itself from the same audio. In the minute or so I ran the test, with my ears I heard probably 5 bird species, Merlin picked up 4 of them, and birdnet-go registered 0 detections.

Given the nature of this test (where the audio is getting picked up by the backyard mic, captured by birdnet-go, streamed to the browser, output over headphones, and picked up again by an iPhone mic), I no longer think the microphone setup on my birdnet box is the biggest limiting factor. (Definitely can still be improved, though.)

It seems like there is plenty of detectable audio data in the system, and for some reason Merlin is currently able to extract more detections from it. My understanding is that Merlin and birdnet-PI/birdnet-go use the same underlying model, which leads to me think this is something about how I have birdnet-go configured that is keeping the detection count so low.

Is there a way to roughly approximate the threshold/settings that Merlin is using in birdnet-go, to lessen the detection gap I see between the two? Or is there something fundamentally different about how the analysis is happening that may be the source of the difference?

[tphakala]

My understanding is that Merlin and birdnet-PI/birdnet-go use the same underlying model

This is incorrect, Merlin has their own model which unfortunately is not open. It seems that Merlin model is being actively developed while BirdNET has not received any updates in more than 2 years now.

[farski]

Ah ok, bummer, but good to know. I will adjust settings and see if I can close the gap through trial-and-error.

[tphakala]

It is possible to improve BirdNET performance by training a custom classifier. I have done so for our native species. I’ve been planning to make improved classifier and training data available, but it is currently in a messy state, and I haven’t had time to work on it in quite a while.

I also reached out to the Merlin team to ask if they would consider open-sourcing their model, but I haven’t heard back, so I assume it’s not likely to happen.