We can do “sound pressure wave to phoneme stream” pretty darn well, and in way that generalizes to anything you have a phonetic mapping for. (cf DeepSpeech, etc

We can do “sound pressure wave to phoneme stream” pretty darn well, and in way that generalizes to anything you have a phonetic mapping for. (cf DeepSpeech, etc) Going from text stream of phonemes (“mmaaaayyynaammeezzzbahhhb”) to text stream of sequence of words (“my name is bob”)is vastly more limited. One of them is speech recognition, the other is language modeling. language comprehension? From experience, it’s cheaper and faster to get results by making a new human :)