I've found the second best strategy is that every time I save a boot image for the embedded SBC, I also save a boot image for the computer it was developed on, so if I need to make any changes I have a known working development environment with the exact version of every library in use, as well as Python itself.
I've found the best strategy is to never use Python (or Node.js, for that matter) on offline embedded systems.
Anyway, if anyone knows any good voice recognition or TTS subsystems that don't use Python, please let me know.