Registering and logging in removes this ad.
Registering and logging in removes this ad.
ZugZwang and Speech Recognition
Interesting story here
http://www.extremetech.com/article2/0,1697,1915529...
about a guy building a chess player, Deep Blitz, based on a pair of AMD 64bit CPUs (Opterons not FX's, for technical reasons) to rival IBM's old Deep Blue.
Reading the article you realize that building a digital chess player and developing SR are formally similar. To quote:
"What is often overlooked in discussions about Deep Blue is that . . . that order of magnitude improvements could be achieved through software/hardware synergy. . . In a larger context, however, solving the problem of developing a program to play chess through a brute force approach shows that computers do not understand chess in the way humans do: They cannot apply general principles taking situational variables into account. Although Deep Blue was a triumph for artificial intelligence in one sense, there's nothing intelligent about brute force calculation."
"Despite its vastly inferior brute force, the Deep Blitz machine could already be a match for Deep Blue because of improvements in chess software. Deep Fritz [i.e., the chess software engine running on Deep Blitz] is able to evaluate lines of play to a similar depth because it successfully narrows its search only to the strongest lines of play."
Don't worry: The machine match never took place, partly because Deep Blue is so old and venerated that IBM wouldn't risk taking it out of mothballs for a potential loss, just as it refused Kasparaov a rematch once it had won.
Deep Blue didn't understand abstract concepts that human chess players use, like ZugZwangs, pins, gambits, etc. Similarly, SR at its fundamental level doesn't understand any part of human speech. However, like a chess player, SR depends partly on brute force calculations (like Deep Blue) but also on intelligence built into subsequent versions (like DNS or VV) that shortcut the brute force approach.
Perhaps the next SR breakthru will derive from whoever is best able to seamlessly integrate more "understanding" into SR algorithms than currently is the case.
Bruce



dear friends , iam
dear friends ,
iam working on continous word speech recognition ,i had already done Isolated word recognition with 95% accuracy . I need help for the continous word recognition to build Language models . please help me out
Regards ,
Murugan.B