<<General problem with them all is they are fixed, meaning the offending sample
is presented, its filtered out and that small section is applied for the rest
of the recording.>>
That's not really how noise reduction algorithms work. The standard model
divides the spectrum into a thousand or more bands & places a threshold
dependent gate in each band. When properly set the threshold should be above
the residual noise but below any desirable audio. Thus, in the presence of
good, well recorded signal the gates are all open & passing full bandwidth
audio, which masks the presumably lower level noise. A band with noise but no
desired frequencies is gated, or turned down in volume. Thus the algorithm is
constantly adapting to the changing conditions of the audio signal. This CAN
work extremely well where there is fairly low level noise & a healthy
separation in volume between that noise & the signal being recorded. When noise
& signal are anywhere close to one another the process breaks down & you can't
really save the recording, as the thresholds can no longer distinguish noise
from signal.
<< Whats really needed is a temporal EQ where EQ is present at a specific
granularity, every second, minute, or right down to milliseconds and blended
between 'temporal EQ frames'. Tracking / learning EQ with noise reduction -
except Ive not found one yet :( Even EQ plug'ins applied on a track affect that
track while its active, imagine if you could blend frames of EQ over the whole
track adjusting for situations where e.g traffic volume increases or decreases.
Or maybe such a thing exists?>>
Every automatable EQ fits this description, although you run into audible
artifacts when automating filter sweeps very quickly. Several seconds, rather
than milliseconds, is a preferable time frame for moving a high pass filter to
cut occasional instances of offending traffic rumble, though often an ultra
quick move to catch a wind gust can be made to not sound too unnatural.
Scott Fraser
|