Вот тонкости
прошу прощения что не на русском
FFT Size
The FFT Size parameter causes the most drastic changes in quality. It determines the number of individual frequency bands that are analyzed. The noise in each frequency band is treated separately, so the more bands you have, the finer frequency detail you get in removing noise.
For example, if there’s a 120Hz hum, but not many frequency bands, frequencies from 80Hz on up to 160Hz may be affected. With more bands, there’s less spacing between bands, so the actual noise can be detected and removed with more precision.
However, with too many bands, time slurring occurs, which can make the result sound reverberant or “echoey” (with pre- and `post-echoes`). So the tradeoff is frequency resolution vs. time resolution, with lower FFT sizes giving better time resolution and higher FFT sizes giving better frequency resolution.
Good settings for FFT Size range from 4096 to 12000.
Remove Noise/Keep Only Noise:
For normal operation, choose Remove Noise.
If for some reason you want to extract the noise for other purposes, choose Keep Only Noise. All the audio will be removed, leaving only noise.
Reduce by
Lowering this value (measured in dB) may help reduce some bubbly background effects. Settings between 5dB and 100dB are recommended.
Precision Factor
This value affects distortions in amplitude.
With values of 3 or less, the FFT is performed in giant blocks that aren’t very continuous between the blocks. This means that after each block is processed, there can be a drop or spike in volume at the interval between blocks.
Values of 5 and up work best. Beyond 10 or so, there is no noticeable change in `quality-only` in the time it takes to compute. Try using 5 or 7, as odd numbers are best for symmetric properties.
Smoothing Amount
Smoothing Amount takes into account the standard deviation, or variance, of the noise signal at each band. Bands that vary greatly when analyzed (such as white noise) will be smoothed differently than constant bands (like a 60 cycle hum).
Generally, increasing the smoothing amount (up to 2 or so) will reduce the “burbly” background artifacts at the expense of raising the overall background broadband noise level.
Transition Width
This setting determines how sharp the division is between what is considered noise and what should be kept. For example, with a Transition Width of zero, a sharp, noise `gate-type` curve is applied to each frequency band. If the audio in the band is just above the threshold, it stays; if it’s just below, it’s truncated to silence.
Conversely, you can specify a range over which the audio will fade to silence based upon the input level. For example, with a transition width of 10dB, and a cutoff point (scanned noise level for the particular band) of -60dB, then audio at -60dB would stay the same, at -62dB it would be reduced some (to about -64dB), and so on until audio at -70dB would be removed entirely.
Again, if the width is zero, then audio just below -60dB is entirely removed, while audio just above it would remain untouched. Negative widths simply go about the other side, so in the above example, a -10dB width would have ranged from -60dB to -50dB.
