Availability of corpora for other languages.

If you care about much less resourced languages than English or the big ones, then you can generally get sufficient text to do interesting stuff, but working with speech becomes much more difficult due to very limited quantity of decent quality data.