[Peter Larsen[_3_]]

bob wrote:

I recorded a live show where there are several instances of mic
feedbacks. I want to salvage as much as possible from those instances.

I have several audio/wave editors (adobe audition, soundbooth,
audacity, goldwave) at my disposal.

Most audio editors have notch filters, but to complicate things, the
mic feedback changes frequency slightly as it gets louder.

Is there a filter that automatically adjust to the frequency change
and remove the feedback?

Probably not, this needs an operator. At a glance - using A3 - I would cut
up in segments and use FFT analysis and FFT eq. It may be possible to define
a few categories. It is possible to define very narrow filters, example for
50 Hz hum: 45 Hz 0 dB, 46 Hz 0 dB, 47 Hz 0 dB 48 Hz 0 dB, 49 Hz -36 dB, 50
Hz -36 dB, 51 Hz -36 dB, 52 Hz 0 dB, 53 Hz 0 dB, 54 Hz 0 dB, 55 Hz 0 dB. The
objective of the extra points is to tell the algorith that you want a sharp
corner.

The FFT equalizer is well suited for removing what you do not want because
it is phase linear. It is less well suited to compensate for minimum phase
deviations such as those caused by transducers. This is not a problem to
solve with the notch filter function. You must use 32 bit audio for optimum
sound quality, especially as it may end up being a multistage process. It
may be that you end up with the filters sharp enough, but have to do the
filtering twice.

Use splines, use dB display. Note: the software is broken by design, so you
MUST define filters in the actually used samplerate, I do not know whether
they fixed it in the new version. It reportedly doesn't have open append,
which I use a lot, so I skipped it. If you divide this in three piles of
numbered segments it may get batchable. Use - ta da - open append to
reassemble the audio, multiselect works, but select from last to first.

There is also frequency display editing in A3, it effectively does about the
same and is perhaps better suited for the actual problem, but you have
better control over the process by the above suggested method.

Kind regards

Peter Larsen