Identifying offending noise by using spectral frequency display

[Battle Vaughan]

Nice one, Jim, want to share how you did it? :)

[Jim Andrada]

Pretty much the way Alexy suggested - using the Deconstruct module in Izotope 5 Advanced.

Honestly I had never used Deconstruct before so I thought it would be a good chance to learn it and I started working in my typical scientific way - in other words playing around with parameters with no clear idea of what they meant, just tweaking and listening to see if things got better or worse. Made a mess and had to start over half a dozen times (or more) but I kept at it and little by little figured it out and arrived at settings that seem to have done the trick

Now comes the hard part - working back from the result to an effective process that will get me there more efficiently.

[Battle Vaughan]

Nice work, nice tool, also. Worth the considerable investment, I expect! Thanks for sharing.

[Jim Andrada]

I think Izotope has been one of my most used packages for the last few years. The only thing I record is classical concerts and it's been enormously helpful in getting out the inevitable coughs and sneezes and whatever. I had a string quartet recording a few years back where the cellist bumped the cello against a chair - horrible sound. I was able to repair it with Izotope. Had another one where one of the French Horns came in early and I was able to hide it pretty well.

I was pleasantly surprised at how well Deconstruct worked - in fact I ran it over the whole clip and I think it cleaned up a lot. I might go back and see how it does with the bumped cello. I'm definitely a fan of the program even though it is a bit pricey.

[Kathy Smith]

The bump is gone but I feel like the vocal is way too distorted in this case. Hearing the distortion of the voice I think I would opt out for something in between minimizing the bump and not completely distorting the voice. Thanks!

[Alexey Lukin]

With Deconstruct, you can always mix the dry and wet signals to your taste: they are phase-coherent.

[Greg Miller]

Sorry to chime in so late, I've been out of town and away from my audio machine for quite a while.

Here's a quick fix. I could probably refine this a bit if necessary. How does this sound to you?

[Kathy Smith]

Greg, that sounds great. What did you use?

[Greg Miller]

Kathy, I'm glad that sounds OK to you.

I used several passes of [adaptive] noise reduction, after choosing the areas for the noise samples very carefully.

In general, I convert the sampling frequency to 96 kHz before processing. That provides more data points for the math to work its magic. (And in this case I converted it to mono, since both channels were identical.)

Then I slow down the playback frequency to 48 or even 32 kHz, so I can hear very precisely when various sounds start and stop. That allows me to accurately choose what parts of the file to use for my noise samples. I don't want to inadvertently include some voice in the noise sample.

In the case of your file, I heard obvious differences in the noise characteristics from one spot to another, so I processed each syllable separately, and then blended all the parts back together by cross-fading them.

I still hear some NR artifacts, especially when listening on headphones. And the timbre of the NRed voice is different from the timbre of the raw voice. (No easy way to avoid that, given the random frequency distribution of the "bang.") But it lasts less than 1/2 second, so hopefully most people won't notice it (especially if they're not looking for it).

As I said, it might be possible to make it a bit better with a lot more work. I posted the "15 minute" version just as an illustration.

[Kathy Smith]

Originally Posted by Greg Miller

Kathy, I'm glad that sounds OK to you.

I used several passes of [adaptive] noise reduction, after choosing the areas for the noise samples very carefully.

In general, I convert the sampling frequency to 96 kHz before processing. That provides more data points for the math to work its magic. (And in this case I converted it to mono, since both channels were identical.)

Then I slow down the playback frequency to 48 or even 32 kHz, so I can hear very precisely when various sounds start and stop. That allows me to accurately choose what parts of the file to use for my noise samples. I don't want to inadvertently include some voice in the noise sample.

In the case of your file, I heard obvious differences in the noise characteristics from one spot to another, so I processed each syllable separately, and then blended all the parts back together by cross-fading them.

I still hear some NR artifacts, especially when listening on headphones. And the timbre of the NRed voice is different from the timbre of the raw voice. (No easy way to avoid that, given the random frequency distribution of the "bang.") But it lasts less than 1/2 second, so hopefully most people won't notice it (especially if they're not looking for it).

As I said, it might be possible to make it a bit better with a lot more work. I posted the "15 minute" version just as an illustration.

That's awesome Greg! Thank you for the explanation. Yes I do hear some NR artifacts too when I listen on headphones but that's so minor that with the music in the background you wouldn't even notice. A question for you. Did you use Izotope RX or something else?

[Greg Miller]

I used Adaptive Noise Reduction in Audition.

[Greg Miller]

Kathy,

Try this version. Maybe you can use this.

(Oops ... I left the sampling frequency at 96 kHz. You can easily downsample it to 48.)

[Don Palomaki]

This may be too simple to work, but are any of the words/phrases that have the noise repeated noise free elsewhere in the audio?. If so perhaps they could be copied and then pasted as replacements for the noisy spots where other processing does not work well.

[Greg Miller]

In my experience, it would be rare for the "replacement" words to be the right pitch, inflection, and pacing. (Of course if you had another take of the same script, that would improve the probability.) There's a slim chance you might get away with it.

[Rick Reineke]

"words/phrases that have the noise repeated"
- I do that frequently and is SOP for dialog editors. Unfortunately documentary type programs rarely have a word or line multiple times in the same tone, but occasionally you get lucky.