Although my time with the OurEcho Challenge has ended, I am still actively working on this project. Just figured I’d document what I’d already done in here :<
For starters, here is my lightning talk on model performance at the GenSea EOY event (time stamp: 33:56-38:46)
ALSO, here’s my instagram page for project MicroBirder
— — — — — — — — — —
Le setup:
A Raspberry Pi 4B is what everything runs on, the TPU is for training, the small black dot near the mouse plug in is a microphone that collects data for classification.
Data preprocessing:
I picked MFCCs as my data form of choice because they were a bit easier to work with for me in terms of cleaning/processing than spectrograms are. Also because that is the data form that BirdNet-Pi uses, and I want my model to be able to work alongside BirdNet as seamlessly as possible. You can see on these MFCC visualizations below that I had to add a bit of padding to get them to all be a uniform size to run through the model.
ROC Curve for vehicle related noise pollution detection:
Hyperparameter tuning:
More updates/photos to come. This is a blog post in the works.