RecognitionService: Delete all in-memory audio before speech end timestamp after we finish transcribing the segment

[soupslurpr]

This will let us transcribe way longer (possibly 37 hours max) without running out of memory as long as speech ends for a few seconds occasionally.