Phil did a really cool thing this week. He gave Sphinx-4 (the speech recognizer written in Java) the ability to detect out-of-grammar utterances. Up until now, Sphinx-4 would always attempt to match what was said to the active grammar, no matter what. For instance, if the active grammar contained only the digits zero through nine, but the speaker said 'sphinx-4', the recognizer would probably return "six four". The application would have no way of determining if this is really what the person said.

This is clearly a problem for many applications. It is important for an application to detect when someone has said something unexpected, something out-of-grammar. So Phil dug in, did some research to learn the best way to deal with out-of-grammar utterances and has implemented a solution that detects out-of-grammar utterances over 90% of the time without unduly affecting accuracy and speed.

Phil documented everything on the RejectionHandling page of the Sphinx-4 Twiki including a great set of references to research on the topic.

Comments:

Post a Comment:
Comments are closed for this entry.

This blog copyright 2010 by plamere