Finds the quietest passage of an audio file, measures the noise floor there in dBFS and gives the matching noise floor setting for ffmpeg's afftdn denoiser, together with the noise level below the signal and what denoising can gain at that level. The file is read in the browser and not uploaded.
Choose an audio file with a pause or a quiet passage
WAV and AIFF are read sample by sample; FLAC, MP3, M4A, Ogg and Opus go through the browser's decoder.
| Quantity | Value | Meaning |
|---|
The file is divided into windows of 50 ms. The quietest run of eight consecutive windows, 0.4 s, gives the noise floor as RMS in dBFS; windows of exact zeros are skipped because digital silence contains no noise.
The nf control of afftdn expects the level of the noise, not a preference. In the measurement for the article on denoising, the best setting landed within 0.4 to 1.6 dB of the real noise floor across seven noise levels, while the noise reduction control changed the result by only 1.13 dB over its whole range. The command therefore sets nf and leaves nr at its default.
The noise below the signal is the RMS of the whole file minus the noise floor. The expected gain is interpolated from the same measurement: 0.15 dB at 40 dB, 1.03 dB at 30 dB, 2.87 dB at 20 dB and 7.42 dB at 6 dB. It describes the test programme of that measurement, not every recording.
A spectrum of the quiet passage shows whether the noise is hiss or hum. afftdn was measured with white noise only; for vinyl and shellac it has presets of its own, noise_type v and s.
Articles on this
Which control of afftdn matters and from which noise level denoising pays off is measured in the article “Denoising With afftdn and anlmdn”.
Sources: ffmpeg Filters Documentation (afftdn) · Audacity: Noise Reduction
The quietest passage has to be pure noise for the value to be right; in a file without a pause it still contains music. The gain figures come from a synthetic test programme with white noise.