For whatever reason it doesn't seem like this is possible in 23.5. If the proxy is imported directly into premiere with no high rez original attached/linked then it'll transcribe normally. It'll transcribe normally when there's an original with a proxy attach, but if the clip is a proxy and the original is offline then transcripts will not generate. Hitting generate after the language/speakers pop up appears just kicks you back to an idle transcription panel.
This limitation/bug is a bit problematic for long form offline workflows where the editors only have access to proxies on the storage, and since this is an offline workflow we can't work with the proxies directly or else there'll be scaling issues during finishing.
The only workaround we can find is bascially generating the transcripts upon ingest which is a bit less than ideal. Its strange since static transcripts can still be generated with these proxy files if in a sequence, but regular transcripts cannot. I'm not sure if this is by design or if this is a bug?