Excellent description. Thanks! Gil
> -----Original Message----- > From: [EMAIL PROTECTED] > [mailto:[EMAIL PROTECTED] Behalf Of Dave Bernard > Sent: Thursday, February 01, 2007 10:49 PM > To: [EMAIL PROTECTED] > Subject: RE: [NF] Speech-To-Text software opinions sought > > > OK, let me jump in here and see if I can help. > > Voice recognition is getting more and more accurate by the day, but it is > important to realize that (1) it's still a very tough problem to solve and > (2) there are parallel efforts attacking the problem. The parallel tracks > are handling (1) an unlimited domain for a single speaker and (2) > a limited > domain for any speaker. > > "Domain" is essentially the body of work that the recognizer can > understand. > By limiting a recognizer to a specific domain, such as "the art > world", you > can accurately handle many types of speakers without training the > engine on > particular speaker voices. This is because you are arming the recognizer > with lots of hints about what is likely to be said, easing accurate > recognition across many types of speakers. > > The flip side is the only way to handle an unlimited domain (that > is, where > there is no way to tell ahead of time what the speaker will utter) is to > have the recognizer learn a particular speaker's voice very, very well. > > These approached are known as speaker-dependent (Dragon) vs. > speaker-independent (Microsoft Speech Services, e.g.) technology. > > An IVR system, such as an automated attendant, needs to be > speaker-independent for obvious reasons. Therefore, the domain it supports > is very limited, often "yes", "no", "one", etc. It is also why you are not > yet seeing widespread deployment of airline reservations systems that > understand "Do you have any flights from Atlanta to Chicago next Wednesday > night?" It's much more difficult to do this reliably in a > speaker-independent way. > > And remember this, too: voice recognition (converting sounds to > text) is not > the same as natural language understanding (ascertaining the meaning of > text). Both are large problems that are still in the early stages of > perfecting. > > Michael Madigan wrote: > > All the automated phone systems use it, so someone's > > got to be doing something right. > > > > > Good point. I've used some phone systems where the electronic operator > was accurate about 90% of the time. > > > -- > Michael J. Babcock, MCP > > > [excessive quoting removed by server] _______________________________________________ Post Messages to: ProFox@leafe.com Subscription Maintenance: http://leafe.com/mailman/listinfo/profox OT-free version of this list: http://leafe.com/mailman/listinfo/profoxtech ** All postings, unless explicitly stated otherwise, are the opinions of the author, and do not constitute legal or medical advice. This statement is added to the messages for those lawyers who are too stupid to see the obvious.
