I've been horsing around with noise reduction software and as an experiment,
tried removing ambient room noise from a recording made in a hotel lounge
(sound off homemade video) in a big open atrium, by sampling different
levels of audience chatter etc. I've created a midi orchestral arrangement
around a particular performance. This singer did a super job on this song,
but with a somewhat sparse, though talented band. The overall effect is
pretty good but the remaining audience noise is a bit distracting. The
particular software I'm using is Wave Repair.

I find I can get rid of some of the noise this way without killing the music
too badly, but what's left behind sounds like "ghost talk". It has this
almost eerie quality that comes and goes.

What I've noticed is that as you get more aggressive, the first thing that
really suffers is the piano. The singer, bass and the midi drums stay far
more intact. Is that because it has complex harmonics some of which
corresponds to the same frequencies as the room noise?

Any other suggestions for this kind of effort?

"Doc" wrote in
nk.net:

I find I can get rid of some of the noise this way without killing the
music too badly, but what's left behind sounds like "ghost talk". It
has this almost eerie quality that comes and goes.

What you're hearing is a powerful NR with a threshold in the middle of the
conversation levels. Audience noise below that level is muted and above
that level it suddenly appears, but only in those frequencies above the
threshold.

What I've noticed is that as you get more aggressive, the first thing
that really suffers is the piano. The singer, bass and the midi drums
stay far more intact. Is that because it has complex harmonics some
of which corresponds to the same frequencies as the room noise?

The piano has a huge dynamic range. The impact of the hammers is very
loud, but the decay quickly drops below your noise threshold, killing the
sustained tone. The singer and the bass are more sustained. The drums are
simply louder than the noise threshold.

Any other suggestions for this kind of effort?

Rather than one all-inclusive pass, try several lighter passes, reducing
the amount of noise you take out in each pass. Lower both the threshold
and the amount of noise reduction if that's possible.

If you do it right, you'll end up with the background voices still audible
all the time but less intrusive. If your software has a frequency response
modifier, remove more bass noise than treble. The hiss is not as
objectionable as the bumps and rumbles of LF.

You won't be able to remove all the audience noise. To do so would remove
a significant portion of the real audio signal. Your NR software is not
smart enough to distinguish the singer from the guy in the front row except
by volume.

"Doc" wrote in
nk.net:

I find I can get rid of some of the noise this way without killing the
music too badly, but what's left behind sounds like "ghost talk". It
has this almost eerie quality that comes and goes.

What you're hearing is a powerful NR with a threshold in the middle of the
conversation levels. Audience noise below that level is muted and above
that level it suddenly appears, but only in those frequencies above the
threshold.

What I've noticed is that as you get more aggressive, the first thing
that really suffers is the piano. The singer, bass and the midi drums
stay far more intact. Is that because it has complex harmonics some
of which corresponds to the same frequencies as the room noise?

The piano has a huge dynamic range. The impact of the hammers is very
loud, but the decay quickly drops below your noise threshold, killing the
sustained tone. The singer and the bass are more sustained. The drums are
simply louder than the noise threshold.

Any other suggestions for this kind of effort?

Rather than one all-inclusive pass, try several lighter passes, reducing
the amount of noise you take out in each pass. Lower both the threshold
and the amount of noise reduction if that's possible.

If you do it right, you'll end up with the background voices still audible
all the time but less intrusive. If your software has a frequency response
modifier, remove more bass noise than treble. The hiss is not as
objectionable as the bumps and rumbles of LF.

You won't be able to remove all the audience noise. To do so would remove
a significant portion of the real audio signal. Your NR software is not
smart enough to distinguish the singer from the guy in the front row except
by volume.