Bird sound ID apps worldwide: comparisons, tricks (1 Viewer)

[jurek]

Spring arrived, birds are singing, and I again play with Cornell Merlin ID app.

In 2 x 10 minutes of listening in a broad-leaved wood in eastern France, Merlin identified 10 species. I identified 14. I and the app both identified 9. I identified additionally: Common Starling, Carrion Crow, Fieldfare, European Goldfinch and European Green Woodpecker. Ebird identified additionally Hawfinch.

Like last year, Merlin app identified Eurasian Blue Tit on the first second far in the background, before I noticed it. Merlin however did not pick a Starling singing very close, for a prolonged time and well visible on the sonogram. I think the app may fail on that Starling sounds are diverse - both high whistles and low clucking sounds.

For the second time, Merlin picked a Hawfinch which I did not hear in the field, could not hear on the recording and could not see on the sonogram produced. Hawfinch is locally common in this habitat, but none was seen or heard that time. I wonder if it is a Merlin ID hallucination? It may be a fake coming from the algorithm training itself on everything together with no comprehension 'there are usually Hawfinches too, so lets write one'.

Any other alternative bird sound ID apps for Europe?

 

[njlarsen]

I wonder if more starling recordings uploaded to ebird checklists would help with the training of Merlin ...
Niels

 

[Brad Hacker]

I have the impression--based on use of the app among California birds/areas I know well--that within the last year, Merlin became more conservative at trying to ID faint sounds and moved toward IDing only sounds with a "high enough" signal/noise ratio; this led to a decrease in the number of IDs and an increase in accuracy. Where I bird locally in California, the app is now very good at IDing sounds and is quite a helpful tool. "Very good" is hard to quantify as the correctness depends on numerous variables, but today, for example, I'd say it scored >90% for singing birds and >50% for single chip notes, although there were so many birds calling and singing at once it's, again, hard to quantify.

 

[Montereyman]

In most cases my iphone is not sensitive enough to pick up birds more than 25 meters from my location. The weakness is with the iphone and not the Merlin app.

 

[Mono]

BirdNET is the main alternative to Merlin.

BirdNET Sound ID – The easiest way to identify birds by sound.

birdnet.cornell.edu

Merlin and BirdNET are both from Cornell University but have been developed by different teams using different algorithms. Merlin is optimised for real time identification whereas BirdNET is designed for batch processing of existing recordings.

There are other apps and devices available that id birds by sound but most of them are running BirdNET, although some might be opaque about that.

 

[jurek]

I wonder if more starling recordings uploaded to ebird checklists would help with the training of Merlin ...

There should be no shortage of Starling recordings... Maybe one could divide Starling recordings into several groups of similar sounds (this can be done using AI, too), and train the algorithm on every group like on a different species? But I have very intuitive understanding how algorithms work.

I have the impression--based on use of the app among California birds/areas I know well--that within the last year, Merlin became more conservative at trying to ID faint sounds and moved toward IDing only sounds with a "high enough" signal/noise ratio; this led to a decrease in the number of IDs and an increase in accuracy.

I am not sure, but indeed, Merlin recently did not identify anything wrong, but much was not identified at all, or after a longer time.

BirdNET is the main alternative to Merlin.

I must check it, last time I tried it did not work offline. Which is a bummer, because it is common to have no internet connection in the field.

 

[Antik]

You can record calls offline with BirdNET but need internet to analyse them.

 

[njlarsen]

There should be no shortage of Starling recordings... Maybe one could divide Starling recordings into several groups of similar sounds (this can be done using AI, too), and train the algorithm on every group like on a different species? But I have very intuitive understanding how algorithms work.

A little more than 4000 in Macaulay, and obviously, not all will be good quality.

I have no idea if your idea of grouping would help. Another hypothesis one could have is that the species has regional dialects and not all are well represented in Macaulay.
Niels

 

[jurek]

I tried Merlin and BirdNET one after another on the same phone, while a Bluetit was singing. Merlin identified it immediately, BirdNET required selecting time etc, and said it cannot hear any bird there. What I am doing wrong?