HackSMU VII / iMasons · ElephantVoices challenge: remove overlapping mechanical noise (airplanes, cars, generators) from elephant field recordings without distorting the elephant calls.

EchoMap is a deterministic, ML-free DSP pipeline that exploits a structural difference rather than a frequency one — elephant rumbles are sustained harmonic bands; mechanical noise is broadband and stationary.

1. Load at native sample rate (48 kHz) — sr=None preserves 10–20 Hz infrasound.
2. HPSS (Harmonic-Percussive Source Separation) — median filters isolate the harmonic component.
3. Noise fingerprint — per-recording, built from the HPSS-harmonic signal's silent gaps so it subtracts cleanly (this was a bug in earlier versions; see process.py docstring).
4. Spectral subtraction on the harmonic signal (step4a).
5. 8–1200 Hz Butterworth bandpass (step4b) — preserves infrasound down to 8 Hz.
6. In-call energy ratio — a proxy metric (not a true SNR; see evaluation below for real ground truth).

Every stage writes real audio + spectrograms (harmonic.wav, gated.wav, cleaned.wav), so the UI stage-toggle shows genuine intermediates.

The in-call/out-of-call energy ratio in the UI is a cleanup proxy and can be misleading on its own. For real ground-truth evaluation we built two off-line harnesses:

5 labeled elephant rumbles × 2 noise types (generator, vehicle) × 4 input SNRs (−10, −5, 0, +5 dB) = 24 mixes. We measure scale-invariant SDR improvement against the clean reference.

Finding: the pipeline is a low-SNR rescue tool. It improves the recordings that are currently unusable, but will damage already-clean audio — do not apply it blanket.

Reproduce: python scripts/eval/build_mixes.py && python scripts/eval/run_eval.py

Generic speech denoisers (RNNoise / Meta Denoiser / Adobe Podcast) high-pass below ~80 Hz because speech has nothing useful there — which would obliterate every elephant fundamental. We simulate that behavior with a 100 Hz Butterworth HPF + aggressive spectral gate (clearly labeled as a simulation).

Finding: ~39 dB advantage in the elephant fundamental band. This is why classical DSP beats a generic ML denoiser for this use case.

Reproduce: python scripts/compare/infrasound_comparison.py

cd backend && python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
cd ..
uvicorn backend.main:app --reload --port 8000

Batch process the dataset:

python -m backend.pipeline.process batch data/recordings data/calls.csv output

• GET /calls — all processed recordings
• GET /calls/{id} — single recording detail
• GET /calls/{id}/audio/{stage} — hpss | gated | clean (real intermediate files)
• GET /calls/{id}/spectrogram/{stage} — same stages
• GET /stats — aggregate metrics
• POST /process — upload a .wav, run pipeline, return results
• POST /query — natural-language query over processed-recordings metrics (Gemini-backed chat; no vision, no classification)

cd frontend
npm install
npm run dev

Set VITE_API_URL=http://localhost:8000 in frontend/.env.

• Not an ML model. No training data, no GPU, no weights.
• Not a call classifier. call_type is passthrough from the labeled CSV column.
• Not a universal denoiser. It's tuned for elephant rumbles in mechanical-noise-dominated recordings; it will hurt already-clean audio (see SI-SDR table).
• Not a fundamental-frequency estimator. At n_fft=4096 / sr=48 kHz, the 8–35 Hz search range has only ~2 FFT bins — previous versions produced a meaningless 2-valued output; this field has been removed.