Inspiration
Many elephant recordings are unusable due to overlapping noise (vehicles, planes, generators). We wanted to recover these recordings and make more data usable for understanding elephant communication.
What it does
Our tool isolates elephant sounds from noisy recordings . It processes the entire dataset (212 instances) and outputs elephant audio, noise audio, and interactive visualizations
How we built it
We created an audio processing pipeline using: spectrogram analysis (STFT) harmonic masking for rumbles frequency-based filtering for other calls NMF for source separation batch processing across categories (vehicle, airplane, generator, background)
Challenges we ran into
Overlapping signals in similar frequency ranges Scaling processing to all 212 instances Handling paths between Colab and local environments
Accomplishments that we're proud of
Processed 100% of the dataset (212/212) Built a scalable and resumable pipeline
What we learned
Signal structure matters—rumbles are easier than broadband calls Combining multiple techniques improves separation Clean pipeline design is critical for scaling Visualization makes results much more usable
What's next
Add AI models for smarter separation Improve non-rumble accuracy Expand into a general wildlife audio analysis tool
Built With
- librosa
- matplotlib
- numpy
- pandas
- python
- scikit-learn
Log in or sign up for Devpost to join the conversation.