... or just have the user repeat some phrases for 5-15 minutes during initial setup? Like you could do with good ol' Microsoft SpeechAPI from the first decade of this century?

The obsession with cloud-based machine learning is creating half of these problems.