AADHAAR-Adaptive Acoustic DSP Help for Avian Audio Recognition
Various interest groups including ecologists require a robust, mass deploy-able, low cost solution to identify bir presence in space-time dimensions. Signal processing and advances in Machine learning have the potential to address above need.
Explore and mature an accurate, automated, and generalized Avian presence classifier based on avian audio signal detection. Classifier aim is binary: Presence or Absence of avian presence, with no species specificity or avian density estimation
3 iterative step exploration: Signal Validation Transformation, Feature Extraction, Machine learning Model Architecture. Performance evaluation using a mixture of confusion matrix and Receiver Operator Characteristic (ROC)- Area Under Curve(AUC).
Achieved an Accuracy of 84% and AUC of 74% on unseen data. Industry best AUC 89%. Thoughtful model design to audio signal and problem definition found beneficial vs a complex neural network model (# of params)