I've had modest success with "MacSpeech Dictate". It uses the same engine as Dragon Naturally Speaking, but has none of the canned read-back training that Dragon requires.
I've been testing it with good quality recordings of our local veterans.(already in digital format on CD) The process I've had most success with is to let MacSpeech Dictate take a run of about 10 minutes, then go back and correct the resulting transcript. Next I re-assign the transcript as the "learning" text that MacSpeech Dictate uses as it re-listens to the same 10 minute segment of the veteran recording. After this initial run, the accuraccy climbs from the starting 50-60%, up to about 80-85% ... still not perfect, but *good enough* to generate text for indexing and keyword searching. If the transcript is going to be exposed/public facing rather than just machine use, then my next step would be to assign the audio file and MacSpeech Dictate transcript to a volunteer/student for final edits... but at least the starting transcript for human editing would be well along the way.
