The VOICEVOX software terms and the zunko.jp guideline allow commercial use with credit but do not mention AI training or machine learning on generated audio; the Docker Hub text names other characters with their own terms (No.7 non-commercial only, もち子さん with company-involvement limits, others on separate sites) that were not read. Evidence that would answer it: the individual character terms pages and the README inside the VVM download.
Why this is still open: The general terms link to no per-character terms, and the per-character pages were outside the budget.