Who here has trained BirdNet to enable sound detection of other avian and non-avian species? I'd love to hear from you and would be grateful if you could share details about your methodological approach and software workflow. Did you train BirdNet with custom labeled data on other birds not represented in the model? Anurans? Insects? Does anyone know of any successes with training BirdNet on bats? Many thanks in advance for sharing! Deep gratitude for this community.
23 February 2025 4:07pm
BattyBirdNET has been a useful tool for some, but I think it is fairly limited on species.
We tend to just use the BTO pipeline, as they offer a lot of small mammals and moths as well.
GitHub - rdz-oss/BattyBirdNET-Analyzer: BattyBirdNET analyzer for scientific audio data processing. · GitHub
BattyBirdNET analyzer for scientific audio data processing. - rdz-oss/BattyBirdNET-Analyzer
We also donate some money to the acoustic pipeline from our sales, so I have a vested interest!
There are others who have made a lot of cool add-ons to birdnet, but fundamentally you may have an issue when it comes to ultrasonic.
I would be interested to hear more!
24 February 2025 4:20am
Hi Danielle,
I have had success in training custom classifiers using BirdNET for Southeast Asian primates, and I'm working to get this research published soon-ish. My main advice would be to have a robust subset of non-target sounds (you can do this by adding a "NOISE" folder), ESPECIALLY from audio data that is from your study site of interest. Any vocalization toward the lower frequencies was more likely to potentially assign things like insect buzzing as detections. I worked iteratively assessing where the model was having issues with false detections. The model was also not performing great as a single-species classifier (even with non-target sounds), but really benefited from the addition of other species specifically occupying a similar frequency range.
BirdNET is great at offering tools for assessing model performance with the "SEGMENTS" feature. This allows you to get a probability cut-off with the confidence score (which will vary for each species), where you can see at what score there is a decent certainty of getting a correct detection. I did do some playing around with the parameters, but ultimately found that most of the default settings performed the best without overfitting.
As far as getting examples of your species of interest, a lot of the free pattern matching tools will suffice with just one example of your target vocalization. I used the Arbimon platform for this purpose. The success you will have from this will ultimately depend on how rare your species of interest is, and how complex their vocalization is. Overall though, I found it much more preferable to parse through a lot of false detections like this than manually scan thousands of hours of recordings. If you have any other questions, feel free to reach out!
24 February 2025 6:00am
Hi! I’ve used BirdNet to train a model for detecting dugong vocalizations. We categorized the data into five distinct dugong call types and added an additional class for background noise (non-dugong sounds). I implemented a custom classifier. From my experience, I recommend carefully curating the training dataset by selecting only medium to high-quality recordings. Additionally, I ensured that only complete calls that represented the fundamental structure of each vocalization were included. This approach helped improve model accuracy and robustness.
25 February 2025 6:00pm
I'm working on training recognizers for anurans. I'm trying to create a Whombat + BirdNet workflow to help me test my model, specifically for false negatives. The segment and review features within the BirdNet GUI are great tools for getting an initial glance at the performance of your model (it has helped me determine which types of sounds are creating false positives, as well as which sites my recognizer is struggling with). I typically work with rare and imperiled species, so it's important to me that I don't have a high rate of false negatives. When you use the review tool it sorts all of the segments you reviewed into two folders, which you can then feed directly back into the model. I copied all of the segments from the negative folder and pasted them into my background noise folder and just doing that once did an incredible job at reducing false positives.
I have found the GUI to be extremely easy to use and it runs pretty fast as well. I think my biggest complaint so far is a lack of documentation about all of the parameters which can be changed- the developers recommend sticking to the defaults for most cases, but without understanding what each of the parameters are I don't know if I have a case which warrants changing them (but there is an autotune feature which will help you search for the best parameter values). The tool is very much under active development, so additional documentation may be added in the future!
As far as data workflows go, I have an annotated dataset that other researchers before me have tried to use (unsuccessfully) to create a recognizer using other tools. My first iteration was about as effective as their best performing model, so I have continued to develop it and so far it is the best classifier that exists for the species. I am now looking at moving on to other species (once I perfect this workflow, of course) and my approach is going to be to sample at sites with known high densities first to get enough training data (I work with rare and imperiled species, so unfortunately there often isn't a robust dataset available for me to use), build the recognizer, then implement it on data from sites with lower or unknown densities.
Ryan Smith