I know this might be a bit weird at first, but if you ever created a vector graphic, there is one simple principle:
Nothing is set into stone. This means, everything is just a setting/value, everything can be restored to default, every change can just be reset or altered. It is purely digital. Add effect (for a selection), remove it. Audacity is more analog, add a fadeout > make the change for good. If you save the file and open it again, that’s what you have now. No going back to the original. That’s what needs to change in my opinion. EVERY CHANGE must be saved additionally to the original and therefore be made possible to remove in order to restore the original version of that specific segment.
If I may (as a visual designer) point out, that the overall look of the icons is similar to VLC. Even tough it is highly functional, for most eyes, it is something you don’t want to look at. Instead the new colored audio waves look very fresh and inviting to click on and experiment.
It’s more than just icons, it’s everything in the menu (position, size, color, alignment, …). But for starters, using Google Font icons could be an idea: https://fonts.google.com/icons
And if you are really open to it, I’ll design you a new app layout…?
I do think switching away from it and towards using icons from a single family was worthwhile. I also don’t find Google’s stuff particularly refined anyway – it’s like the Arial of icon sets IMHO: Good for a wide variety of cases, but outdone in almost every individual case by something else, plus it’s got some weird edges to it.
I would encourage you to actually try a redesign for Audacity 4. There’s two ways you can try it, one would be unconstrained by implementability (ie: ignore what’s possible in the current framework and do what you think would be best for users), another one would be by actually trying to implement it within the constraints of the current system. Both approaches are valuable, with the latter being more likely to be adopted wholesale and the former being more likely to get learnings extracted out of.