Inspiration
ElephantVoices researchers have hours of field audio from Amboseli where elephant rumbles at such a low frequency, near the edge of human hearing, along with being buried under generators, vehicles, and airplanes. We created a pipeline that preserves what matters to the elephants and not what matters to human voices.
What it does
EchoMap takes raw Amboseli field recordings and runs each one through a 3 stage DSP pipeline that separates elephant harmonics from the mechanical noise, then serves the cleaned audio, before/after spectograms, and recording metrics through a research dashboard.
How we built it
- HPSS (Harmonic-Percussive Source Separation): decomposing based on structure isolates horizontal harmonic bands (rumbles) from percussive noise, fully deterministic, no ML.
Challenges we ran into
- Elephant rumbles are quiet, where vehicles can leak into HPSS's harmonic channels, making the spectral subtraction aggressive. We switched from hard subtracted to a Wiener mask with a residual floor rescued the recordings.
- Infrasound is hard to visualize.
Accomplishments that we're proud of
- Every call in the labeled dataset is now audibly present in the cleaned output.
- Real SI-SIDR evaluation, our pipeline shows the expected low SNR pattern.
- A pipeline a researcher can actually inspect, where every stage writes its own .wav and .png, no black box.
What we learned
- DPS meats ML when the signal has a known geometry and you don't have a labeled training set.
What's next for Elephant
- Extend beyond elephants...
Log in or sign up for Devpost to join the conversation.