Since some messages are static (for a given version of Wikispeech) it makes sense to generate them when installing or upgrading. There are also some messages that can't be synthesised when they're needed, e.g. if there's no connection to Speechoid. Create a maintenance script to synthesise messages.
Figure out what messages that need to be synthesised. Messages may be dynamic, containing parameters, wikitext magic words etc. In these cases you may want to generate a set of messages with different parameters, where these are a finite and reasonable number. In other cases it may not be feasible e.g. if a parameter is a page name.
Messages should to be synthesised in all languages. They may be synthesised in all voices, though it may be acceptable to to just have them in one voice per language if that would turn out to be too much. It may even be preferable to have a distinct voice for messages to contrast them with other utterances.