Denoising works best when the denoiser knows the noise. This tool captures that knowledge as a reusable fingerprint: mark a region that contains ONLY the unwanted noise (room tone, tape hiss, hum, an air conditioner between sentences), and the extractor decodes it and measures the noise floor level in dBFS, the third-octave band spectrum, the spectral flatness (Wiener entropy - near 1 for white noise, low for tonal hum), and mains-hum harmonics at 50 or 60 Hz plus overtones via exact-frequency Goertzel probes. It then classifies the noise (white hiss, pink, low rumble below 120 Hz, 50/60 Hz hum, broadband, or mixed) and derives tuned FFmpeg afftdn settings from the measurement: the noise floor parameter is set from the measured region level (clamped to the filter’s -80..-20 dB range) and the reduction amount from your chosen strength - rather than leaving both at defaults. Three modes: extract-and-denoise renders a cleaned copy of the whole file through the tuned chain (plus a protective high-pass when hum or rumble dominates); extract-only writes the fingerprint as a JSON file for the record; apply-profile loads a previously saved JSON and denoises ANY file with it - the reuse path for batch cleanup of recordings from the same room, device or tape. The JSON carries the classification, measured floor, band levels, hum frequencies and the recommended filter settings, so the profile is auditable, not a black box. The metadata reports the measured floor before and after denoising (quietest 10 percent of 100 ms blocks), so the effect is verifiable. Feed it honest noise-only regions: a region containing speech will be treated as noise and the denoiser will eat the speech.