Wow, interesting topic. Looks like I'm a little late to the party. I will be responding mainly to the original post as well as Jaybird and Ironcross, who I think have brought up really interesting points. First though, I should probably mention something that could prevent this sort of mess in the first place.
I'm happily mixing sounds, like chemicals in a lab with my sound editor, but at one point I go one command too far and insert 1 extra sound and now the hole thing sounds awful, and since I thought it'd sound nice, I saved and closed it.
Now, I'd want an app that I can give it that 1 sound that I didn't want in the source file, and this app will remove that sound from the source.
I've been there more than once, and it isn't fun.
Most editors like Gold Wave, Wave Pad etc. are destructive, and what you have described is precisely the main downside to destructive editing. In destructive editing, applying an effect will make a new temporary copy of the data with the result. Most editors have an undo feature, which reverts back to the previous copy. When you save the file, you are essentially committing to whatever copy of the file that was open at the time, and erasing the temporary copies that got you here. The only way to revert any changes after saving is to start from an earlier saved copy, and if you didn't make one, you'll have to start over I'm afraid. For this reason I really do not recommend destructive editing when working on complex projects. You always run the risk of committing to something too soon and not being able to go back and revise it.
I think for these kinds of things you should use a DAW that is non-destructive. I use Reaper, which is very accessible with the Osara add-on, though it's a bit of a learning curve, and I am not too proficient in using it yet. but I'm getting there.
Reaper, as well as most other DAWs so far as I know, are non-destructive. In non-destructive editing, everything you do happens as the files are played. The audio is read from disc, and processed in realtime. Because of this, you can review and change things to your heart's content and instantly hear the results. When you perform a save operation, you are only saving a file telling the program the steps that got you here (this is often known as a project file). You can edit your project as much as you like. You could, for instance, decide that you don't like a particular reverb, or you don't like a particular sound. No problem. Just take the offending thing out of your project and replace it with something else. When you are really finished tinkering, you can make what is called a render, which is just a single wav or other audio file that reflects all your work. But even after rendering, you can still tinker with your project, and if you find something you like better, just make a new render. The only way to really get screwed over is to lose your project file.
Now because destructive and non-destructive are internally very different, some tasks are more suited for one than the other. And Gold Wave is still easier to use and learn, so I'm not at all suggesting that you shouldn't use it. But for the problem you're facing, non-destructive editing will make your life easier if you can get to grips with an editor that works that way.
Now for actually removing and isolating sounds. I've been experimenting with this for a while now. It's one of those topics that fascinate me. I've found that there are basically two types of sound removal: inverting, and what I call carving.
Jaybird in post 9 explained inverting very well. You basically take the sound you want to remove and invert it. Gold Wave has an invert effect for this purpose. It won't sound any different when you do this, but trust me, it does do something. Now, mix that inverted copy on top of your file containing many sounds. The inverted copy has to be at the exact same time and volume for this to work. If it works though, the sound will be removed without a trace!
This is purely a mathematical binary operation, so it *only* works if the sound you are trying to remove is exactly identical to the sound in the mixed file. Thus, the two have to be exactly the same volume and at exactly the same time. And anything that could change one or both files, such as clipping, distortion, effect differences, lossy file compression (mp3, ogg, wma etc.), or anything else will make the inverting trick less effective, and as a result, remnants of the sound will be left behind.
Now the method I sort of call carving. Ironcross mentioned it when he said that a sound is essentially a sum of sine waves and that you can remove one sound from another by removing the sine waves of the unwanted sound. He is right in saying that it can degrade the quality moderately to severely. However, using Gold Wave's noise reduction, or any general noise reduction tool for that matter, demonstrates nothing. Noise reduction tools rely on static snapshots of a short sample of background noise which they will remove from the file. The less the background noise changes, the better. The more it changes, the worse the result.
For instance, if a sound you want to remove started off at a low pitch and then increased to a high pitch, and you copy this to Gold Wave's clipboard to try to remove it, the noise reduction would see that there are low and high frequencies present in the sound you want to remove, but it would not know when to remove which ones. Thus, any frequency that ever existed in the sound you are trying to remove will be swallowed up, which of course is going to sound terrible. Or, if the sound starts off quiet and gets louder, the noise reduction will tend to gravitate to the loudest parts. It will not know to be gentler during quieter bits.
To have a fair chance of removing sounds that change over time, you need something designed for this purpose, something to carve out the correct frequencies at the correct times and at the correct volumes. Fortunately, a few things which do this exist. The two options I know of and have tried are the kn0ck0ut vst which I will just call knockout, and Izotope RX now has a de-bleed module which can do this too. The former is free, the latter is not.
Knockout basically has two modes: extract the center (useful for vocal extraction), and extract one channel (you put the full mix on one channel and the sounds you want to remove on another, but I can't remember which channel should be which). Of course, the two should be as similar as possible, so they should be in sync, and the same volume. Unlike inversion though, the sounds don't have to be exactly identical for removal to be effective. They just need to be as close as you can get them.
Unfortunately the output of Knockout is mono since it only has two inputs, one for your full mix and one for sounds to remove. I would've preferred the sounds to remove be on the sidechain, then you could process in stereo. Most modern VSTs that take two types of input use the sidechain, but Knockout is old and I don't think it's being developed any longer. With some routing though, you could set two instances to work by sidechaining. I haven't tried this yet, but I presume it would work. The only real issue with Knockout is that it does produce artifacts especially if the sound you're trying to remove is loud. Perhaps newer things will come out in the future that can remove sounds more affectively and with less artifacts.
The de-bleed option in Izotope RX has less artifacts most of the time, though it isn't as aggressive as Knockout. De-bleed isn't a VST however. it's a module in the RX standalone application, which is actually pretty good for accessibility. And there's a fully functional demo, so you can try it if you want.
Few, that's it I think. I hope you've found something helpful in this rambling! Lol
Make more of less, that way you won't make less of more!
If you like what you're reading, please give a thumbs-up.