somebody asked a couple days ago what i had done with Hidden Markov Models (HMM), so this is that list :
first, i did the obligatory feed the HMM the works of some famous person, and then that can be used to generate sentences that would sound like something the person might say. so for kicks, i fed it the presidential debates and had it kick out sentences for each candidate. of course the bush HMM was constantly saying 'september 11' and 'weapons of mass destruction'.
second, i tied it to the /aiTabletOcr article. did this by populating the HMM with the words in the Tablet PC dictionary. so given a starting letter, it knew what letter was most likely to follow next. then i tied it to the stroke recognition that was already being done to make it do word recognition as opposed to character recognition. it sort of worked ... but i didnt spend much time on it.
third, i used one with the /ttSpeech article. fed it the CMUdict list of word pronunciations to have it learn which phonemes were used most in the English language, and in what order. then i added some randomness so that it would speak back gibberish sentences that was not english, but sort of sounded like it. this actually could serve a purpose, such as being the spoken language for the Sims characters.
finally, while revisiting the /noReco article. the published article only works with single words, and can only recognize a small vocabulary. tried to rework it to recognize phonemes instead, and then use the HMM above to try and determine what word was actually being spoken. almost identical to how i tried to use an HMM with a tablet above, but using phonemes instead of characters. the problem was that i never could get phoneme recognition to work very good.