ElephantVoices Denoiser

A signal-processing pipeline that cleans elephant infrasound where off-the-shelf denoisers fail. It parses CSV/XLSX annotations, loads call segments at native sample rate, and extracts noise-only gaps to profile generator hum. The processing chain combines STFT, HPSS-style harmonic/percussive separation, comb masking, and scoring — HPSS informs f0 detection, but reconstruction uses the original magnitude × comb mask + original phase, with noisereduce applied last on the comb-masked waveform. Optional multi-speaker separation is supported. Exposed through a FastAPI service with a React + Vite demo UI. The pipeline has since been used in real bioacoustic research.

01 / 02