The noise removal process you're referring to is based on 'noise
printing': taking a sonic fingerprint of the noise and building
filter(s) to remove it. The success of the process relies very much on
the quality of the noise print used, so it is important to choose a
good segment of the recording that has plenty of the noise but no
other sounds of interest or value.
As others have pointed out, noise printing can be very effective but
it is easy to overdo it, resulting in two types of artefacts. The
first is the 'birdies' effect others have mentioned here, which
results in short whistling sounds that usually accompany transients in
the audio. The second causes a kind of gurgling underwater sound, as
if the sound is being heard through running water.
When used correctly, noise printing is a powerful tool. Often 3dB to
6dB of noise reduction is all that's required to make a perceived
improvement, and one that is far less obtrusive than trying to remove
the noise altogether. It is always worthwhile ABing the processed
version against the unprocessed version to check that you have
actually made an improvement to the intelligibility of the signal.
On this topic, another powerful tool that is becoming widespread is
spectral editing. It's great for removing unwanted sounds of short
duration like keys jingling, car doors closing, chairs squeaking and
so on. By way of example, today I spent a few hours recording the
semi-finals of an operatic aria competition. The MC was using a
stopwatch to time each contestant's performance, resulting in a little
stopwatch 'beep' about a second after the beginning of each piece, and
another about a second before the reverb tail of the room had finished
fading out. That's the kind of problem that spectral editing is good
for...
- Greg Simmons
|