The original language models were only half of a speech recognizer, whose function was to produce coherent text. The other half was the acoustic model, whose function was to ensure the text corresponded to reality. We forget the latter at our peril.

→ View original post on X — @pmddomingos,