Improved BirdNET classifier

[tphakala]

Hello All,

I have been working on improved BirdNET classifier, it adds few new species and improves detection of some existing species and human made noises. Classifier is available at https://github.com/tphakala/birdnet-go-classifiers, currently there is only English label file available.

Instructions on how to use custom classifier with BirdNET-Go is under Installation & Usage section of classifiier repository.

Here is quick list of classes included in custom classifier sofar

Augmented Classes
These species have enhanced detection capabilities compared to the base BirdNET model:

• Certhia familiaris - Eurasian Treecreeper
• Corvus corax - Common Raven
• Corvus cornix - Hooded Crow
• Cyanistes caeruleus - Eurasian Blue Tit
• Dendrocopos major - Great Spotted Woodpecker
• Dog - Dog
• Dryocopus martius - Black Woodpecker
• Engine - Engine
• Garrulus glandarius - Eurasian Jay
• Glaucidium passerinum - Eurasian Pygmy Owl
• Human vocal - Human vocal sounds
• Loxia curvirostra - Common Crossbill
• Parus major - Great Tit
• Picus Canus - Grey-headed Woodpecker
• Pyrrhula pyrrhula - Eurasian Bullfinch
• Siren - Siren (added to improve siren detection and reduce false matches with Eastern Screech-Owl)
• Tetrastes bonasia - Hazel Grouse
• Turdus merula - Eurasian Blackbird

New Species
Species added to extend the base model's capabilities:

• Ovis aries - Sheep
• Vulpes vulpes - Red Fox

[alexbelgium]

Hi, thanks very much for making this available! I've updated the translated labels here using an automated code that either copied the common name from the translated labels if existing, or attempted to fetch it from wikipedia in the correct language. This is not perfect, but already some way in the right direction ;-)

Btw not sure if it causes an issue but you have Picus canus_Grey-headed Woodpecker with a capital C and a a, while the initial file contains Picus canus_Gray-headed Woodpecker with a small c and a e

BirdNET_GLOBAL_6K_V2.4_Labels_af.txt
BirdNET_GLOBAL_6K_V2.4_Labels_ar.txt
BirdNET_GLOBAL_6K_V2.4_Labels_bg.txt
BirdNET_GLOBAL_6K_V2.4_Labels_ca.txt
BirdNET_GLOBAL_6K_V2.4_Labels_cs.txt
BirdNET_GLOBAL_6K_V2.4_Labels_da.txt
BirdNET_GLOBAL_6K_V2.4_Labels_de.txt
BirdNET_GLOBAL_6K_V2.4_Labels_el.txt
BirdNET_GLOBAL_6K_V2.4_Labels_en_uk.txt
BirdNET_GLOBAL_6K_V2.4_Labels_en_us.txt
BirdNET_GLOBAL_6K_V2.4_Labels_es.txt
BirdNET_GLOBAL_6K_V2.4_Labels_et_ee.txt
BirdNET_GLOBAL_6K_V2.4_Labels_fi.txt
BirdNET_GLOBAL_6K_V2.4_Labels_fr.txt
BirdNET_GLOBAL_6K_V2.4_Labels_he.txt
BirdNET_GLOBAL_6K_V2.4_Labels_hi_in.txt
BirdNET_GLOBAL_6K_V2.4_Labels_hr.txt
BirdNET_GLOBAL_6K_V2.4_Labels_hu.txt
BirdNET_GLOBAL_6K_V2.4_Labels_id.txt
BirdNET_GLOBAL_6K_V2.4_Labels_is.txt
BirdNET_GLOBAL_6K_V2.4_Labels_it.txt
BirdNET_GLOBAL_6K_V2.4_Labels_ja.txt
BirdNET_GLOBAL_6K_V2.4_Labels_ko.txt
BirdNET_GLOBAL_6K_V2.4_Labels_lt.txt
BirdNET_GLOBAL_6K_V2.4_Labels_lv_lv.txt
BirdNET_GLOBAL_6K_V2.4_Labels_ml.txt
BirdNET_GLOBAL_6K_V2.4_Labels_nl.txt
BirdNET_GLOBAL_6K_V2.4_Labels_no.txt
BirdNET_GLOBAL_6K_V2.4_Labels_pl.txt
BirdNET_GLOBAL_6K_V2.4_Labels_pt_BR.txt
BirdNET_GLOBAL_6K_V2.4_Labels_pt_PT.txt
BirdNET_GLOBAL_6K_V2.4_Labels_ro.txt
BirdNET_GLOBAL_6K_V2.4_Labels_ru.txt
BirdNET_GLOBAL_6K_V2.4_Labels_sk.txt
BirdNET_GLOBAL_6K_V2.4_Labels_sl.txt
BirdNET_GLOBAL_6K_V2.4_Labels_sr.txt
BirdNET_GLOBAL_6K_V2.4_Labels_sv.txt
BirdNET_GLOBAL_6K_V2.4_Labels_th.txt
BirdNET_GLOBAL_6K_V2.4_Labels_tr.txt
BirdNET_GLOBAL_6K_V2.4_Labels_uk.txt
BirdNET_GLOBAL_6K_V2.4_Labels_vi_vn.txt
BirdNET_GLOBAL_6K_V2.4_Labels_zh.txt

[tphakala]

Btw not sure if it causes an issue but you have Picus canus_Grey-headed Woodpecker with a capital C and a a, while the initial file contains Picus canus_Gray-headed Woodpecker with a small c and a e

Yes this is an issue, thank you for pointing it out!

[Tserppa]

Hi, feeling bit stupid here and can't get it working. Downloaded the 20250909 files and placed them in data/models/ and set up config.yaml accordingly.

I'm getting: "Command execution error: failed to initialize BirdNET: BirdNET: failed to initialize analysis model: cannot load TensorFlow Lite model"

What am I doing wrong?

[tphakala]

Not knowing exact settings you applied, but model and label settings under birdnet section should look like this:

birdnet:
    modelpath: models/BirdNET-Go_classifier_20250909.tflite
    labelpath: models/BirdNET-Go_classifier_20250909_Labels_fi.txt

Also here is label file with finnish species names

BirdNET-Go_classifier_20250909_Labels_fi.txt

[Tserppa]

Matches with my settings. It definitely finds the files but for some reason cannot initialize it. Is there something else I can double check?

birdnet:
    debug: true
    sensitivity: 1.5
    threshold: 0.93
    overlap: 1.5
    longitude: xxxxx
    latitude: xxxxx
    threads: 0
    locale: fi
    rangefilter:
        debug: false
        model: latest
        modelpath: ""
        threshold: 0.018
    modelpath: models/BirdNET-Go_classifier_20250909.tflite
    labelpath: models/BirdNET_Go_classifier_20250909_Labels_fi.txt
    usexnnpack: false

[tphakala]

how did you download files? check that BirdNET-Go_classifier_20250909.tflite is actual tflite file and not HTML, file size should be 60292268 bytes

[Tserppa]

Got it now. Used wget but seems like I don't know how to use github... Thanks!

[tphakala]

I hade same issue! I have a plan to add automatic model download and install in UI soon.

[leccelecce]

Just to check, is this classifier automatically included in the nightly builds? Or if we want to use it do we have to specifically enable?

I looked in the UI but couldn't see anywhere obvious where it tells you the model in use.

[tphakala]

Thanks, I've given this a go. Surprisingly for me it seems to be much less accurate.

1. Distant motorbike sounds / planes going over through the night are now detected as Rock Pigeon, which wasn't happening before

2. No reduction in sirens or just wind noise being detected as Tawny Owl (which was the main improvement I was hoping for)

3. Eurasian Bullfinch and Great Bittern is now being detected though these are definitely false positives

I'm running nightly-20251008-16-gc191be09 and the 0916 model.

I'm going to revert back and confirm things return to normal.

Can you make those false detection audio files available for download so I could improve model with them?

[leccelecce]

Thank you for the fast response! Yes, will collect some together a bit later and share.

On a side note - I set the new model path in my config.yml. I then saw it can also be specified in the UI config pages. I deleted the values out to disable it, clicked Save and restarted, but it seems it doesn't actually update the config file.

Oddly, I can see the config file timestamp does change when I click save, but it's not actually resetting the two fields back to "". Is that expected behaviour?

[leccelecce]

pyrrhula_pyrrhula_78p_20251019T095423Z.wav
pyrrhula_pyrrhula_98p_20251019T095359Z.wav
pyrrhula_pyrrhula_86p_20251019T020736Z.wav

[leccelecce]

And the "great bittern":

botaurus_stellaris_71p_20251019T002400Z.wav
botaurus_stellaris_83p_20251019T043117Z.wav
botaurus_stellaris_86p_20251019T045358Z.wav
botaurus_stellaris_96p_20251019T030203Z.wav
botaurus_stellaris_81p_20251019T031411Z.wav

[jmbaggs]

Can you make those false detection audio files available for download so I could improve model with them?

@tphakala : I also have some false detections happening, like when my neighbor opens their door the noise it makes is sometimes identified as a bird! Would you also like these files?

[francois2k1]

I get a lot of Siren detection in UK city, usually detected as water birds or Owls.
How could we improve the software to have a Siren detection feature based on location (UK Sirens, French Sirens...)
Could it be implemented in the same way as dog backing detection?