Interesting article on using common sense information in speech recognition:

http://www.trnmag.com/Stories/2005/032305/Common_sense_boosts_speech_software_032305.html

Users appear to like their experience better if the result (even if incorrect) is semantically plausible.

Better late than never, eh?

Now that there are more speech bloggers out there, I should probably tell you who I am.  My name is Jen, obviously, and I've been working in speech here at Microsoft for about the past three years.  My primary focus is everything desktop related, but I spend a lot of time out here reading about customer experiences, wants, needs, etc. and will probably focus on that more than anything.  I'd like to see desktop speech in more mainstream use and I'm looking for ways to make that happen.

I ran across this aricle today: http://www.vnunet.com/news/1160755 about the British Computing Society encouraging UK researchers to investigate implementing more human-like behavior in computers.  The article talks some about cognitive processing, invesitgating mental disorders and intelligent robots.

But, surprisingly, speech was not mentioned.  I personally think that speech is the next natural interface with computers.  But does natural equal human-like?  In my opinion, yes.  But not as the current technology stands.  I think that in order for people to use speech widely on desktop systems that they want an interaction that rivals that of speaking to a human - more on the lines of semantic interpretation.  We need to make speech be exciting for a user, to provide some value.  An alternate way of typing is novel, but not that thrilling - especially given that typing is taught to most students in high school or younger.  For persons with RSI or some other disability, I realize that dictation (good dictation and command and control) may be enough.  But, for the other users, they want more - they want to accomplish something that they can't do otherwise, or accomplish something in a faster, more natural way.  Currently, I don't think there's any SR system on the market that allows for this sort of experience.

I think we are coming close in embedded applications, but we are limited by memory.  I've seen academic demos of some really amazing applications that do add value to the end user - online recognition and translation for example.  Things like controlling smart homes are intriguing as well.  I know that when I'm laying on the couch I would love to just be able to say "set temperature to 68 degrees" and have it work.  Save me some work and I'll use SR for everything.  That's the message I'm getting from most non-impaired users.

So, I'm a linguist at heart.  Give me a sentence and I'll give you a syntactic and/or sematic breakdown for your viewing pleasure.  So, Fil sent around this article today and now I'm excited about linguistic search engines and corpora and everything.  I need a sentence to parse.

Seriously though, I often think about the relation to SR engines and syntactic / semantic grammars.  I don't know of any engines (aside from research ones) that actually consider straight syntactical information in their evaluation.  What people really want is a semantic engine.  They want to speak to the computer like they talk to you or I.  Who wouldn't?

If you like things like this as much as me (which is doubtful) then check out these fun places.  I won't tell.

Original article: http://www.economist.com/science/displayStory.cfm?story_id=3576374

Linguist's search engine: http://lse.umiacs.umd.edu:8080/ - so you can look for things like <noun phrase> <verb phrase> instead of specific words

Language Log: http://itre.cis.upenn.edu/~myl/languagelog/ - a blog for linguistics!  My life is complete.