I hope this is ok to post here. I figured there was probably a lot of overlap with iNat and Merlin users and there doesn’t seem to be a forum for the Merlin app.
I’m pretty bad identifying bird calls but am trying to learn and finding the Merlin app helpful. I’d like to get better with it before the big spring migration starts. For many birds the ids seem accurate. Common yard birds where I can see them calling and the app is picking them all up and identifying them properly in real time… is pretty cool.
The problem I’m having is when it identifies something more unusual that I’m not seeing or hearing. Was it a quick flyover, very faint, outside my hearing range? I know I can tap on the bird shown in the playback menu and it jumps to the section of the recording where it was first heard. I usually don’t hear anything significant when I play it back though. I don’t know when the call it identified actually starts and stops… is there lead time before the call? It doesn’t highlight the spectrogram where it identified the call or seem to have a way to jump to any later calls from that bird in the recording. I saw you can tap and drag to highlight a section of the recording and it will run another identification on it, but when I do that it never seems to identify that bird again.
The documentation in the app and their website is very basic. So I’m wondering if anyone here had more knowledge about using it. I’ve only ever used a recording from Merlin once to submit to iNat as an audio observation, but I’d like to do more if I can actually figure these things out to make quality audio observations.
I use Merlin some; my experience is this: don’t totally rely on the app; it’s often correct, but not always. At my home it semi-regularly “identifies” birds that I know don’t normally, if ever, occur here. Try to listen to the bird yourself, the closer the better. Distant calls or lots of birds singing at once are scenarios that can lead to misidentification. Chip calls are more likely to be misidentified than songs. When playing back your recording in the app, you may have trouble hearing the fainter sounds. You can try sending the recording to your computer and using a software such as Audacity to normalize the recording; ideally, recordings uploaded to iNat or eBird should always be normalized. You should then be able to hear the calls better, and I suppose you could even play it back and let Merlin re-identify it to know exactly which calls are being marked as which species. Personally, I generally don’t take the time to normalize recordings for iNat, but for eBird I do.
Merlin very frequently “detects” birds that are not actually there. As discussed in many threads, you should not use auto-recognizers (including iNat’s CV) to ID beyond what you can verify (with assistance from knowledgeable human identifiers, if needed).
Merlin is a great tool, but overreliance is causing a lot of data quality issues that are frustrating eBird reviewers in particular. The good thing about Merlin is it creates an audio file that can be verified by others if you are unable to make a field ID. And then those audio files can be edited (some relevant info in this thread) and uploaded to citizen science platforms like iNat and eBird.
just speaking generally, a lot of times, unexpected IDs are related to the training data and / or the recording itself. (higher quality recordings in general will be more likely to provide good ids.) just for example, if you have a little bit of handling noise during recording, that could be interpreted as rattling bird calls. or wind or other background noises that muddle the recording of a sound may make it hard to match that sound accurately.
i wouldn’t expect there to be much if any lead time. folks who are annotating the sounds that go into the training set should be making very tight boxes around the relevant parts of spectrograms. see: https://merlinvision.macaulaylibrary.org.
generally, the microphone on your mobile device won’t even record much beyond normal human hearing range, if at all. so i doubt this would be the source of odd ids.
…
you could also try using BirdNET. i believe the way that works is based on the last n seconds of a given recording. so ids will drop off a few seconds after they are no longer detected, and you’ll get indications of the “certainty” of the ids.
Yeah, I’m not submitting any observations here or on E-bird based on Merlin suggestions only. The only one I submitted I included the clipped and normalized recording so that others could hear it too and the Merlin suggestion on that one was actually for the wrong owl. I’m mainly using it to learn and provide recordings for more skilled birders to help me identify the less common birds.
I’d tried to re-run Merlin id on some audio recordings after normalizing the audio on my computer and got completely different suggestions which was weird. So I know it definitely has flaws.
When I was just testing the app again I thought maybe they had updated it, because it was working differently. I was able to tap on a bird in the list under the saved recording and it jumped to each instance of the bird with each subsequent tap. That’s how I thought it should work initially, but it never had for me. It also highlighted each bird in the list as it appeared on a recorded playback just like when it was initially recording them. Still though when I selected a large area of the spectrogram to rerun the id it never identifies the same bird there again. Strange thing is when I tried playing back some older saved recordings the features mentioned above didn’t work at all on them. It also seems to have issues freezing up on longer recordings or if I tried to switch to another app during use, then the recordings seem corrupted or something and won’t load for me to play them back.
I really wish it would highlight the exact part of the spectrogram that it is using for the id. That way I could see what section to try and clip and normalize to listen to more closely. Or, maybe just see that it made that wild misidentification based on obvious handling noise or the wind chime, etc.
Is there a way to easily clip sounds within Merlin so they can be uploaded to iNat and the IDs verified? I just re-downloaded this and have been really enjoying using it, but i am not good with birds and i know just like the iNat ID i can’t really rely on it to make reliable observations of these species without additional info. But i am guessing a merlin ID plus iNat verification by other people would be a lot better.
There is none within the app that I am aware of. And just to clarify my previous comments, I really appreciate the app, I just use it with caution. If Merlin can hear the bird in question well and repeatedly identifies it as the same species, it’s probably right. Distant, obscured by background noise, etc.? Don’t rely on it.
Absolutely agree. I would emphasize again that Merlin is a great tool for preliminary IDs and I’ve found it generally more accurate than people in this forum seem to indicate. However, it does have its flaws; I would recommend using it as a starting point and then looking for the birds it IDs (if it seems reasonable for the location).
it’s certainly more accurate than my plant-biased brain at identifying birds :D I wish they’d make a crossover with iNat though, or even a better crossover with eBird, i tried cross posting a few things to eBird and that didn’t seem a very efficient process either.
I was playing around with the app some more and figured out a couple more things. I was looking at some of my recordings from last fall that I’d saved because they supposedly had “rare” birds in them that I wanted to look into more. The jumping ahead to the bird call in the recording and highlighting birds on playback wasn’t working like I mentioned in an earlier post. When I hit the edit button though where you can change the filename, location, and date… I noticed a new setting “update suggestions”. I tapped that and it took a few minutes but seemed to be reanalyzing the file. The suggestions changed and those rare birds were no longer in the list! No snow bunting in my back yard! The jump to call and highlighting the bird during playback are now also working on those older recordings.
I tried doing the same thing on some of my more recent recordings that weren’t working correctly, but the option wasn’t in the edit menu for them. “Update suggestions” also is now no longer in the menu for those older recordings. It must be that it tags a file with whatever version of the dataset it used to make the current suggestions and if there is a newer dataset available it gives you the option to rerun the analysis.
The only way I could get the app to redo the analysis on the newer glitchy files was to export the wav file, delete it in the app, reimport it, set the location and date, and then let it recreate the suggestions. It still seems to be unable to properly handle larger files though. I had one that I let run for 20 minutes and trying to open that one just bumps me back to the main menu with an error message.