Speaker-independent speech recognition?

I've got a very specific speech recognition application in mind, and
I'm looking for a reference that will indicate if it's feasible. I
want to recognize just one magic word, which is a very well-solved
problem with high accuracy if we were talking about a boom mike and a
silent environment. The difficulty is that there may be lots of other
noises in the background, other people saying things, etc.

The application is something like a telematics device where you get
its attention by saying "Computer...", except that the word in
question can be assumed to be a unique word nobody would ever use for
any other purpose. However the specifics of this application are
something along the lines of:

- If the computer doesn't recognize that you want its attention, a
ninja will beat you to death with a frozen muskrat, and
- If the computer hears your dog barking and thinks it was you trying
to get its attention, you'll be charged $1,000 for the CPU time.

Is there an article someone can reference for me that will give some
feel for the best I can expect from today's technology? Ideally, some
information on the upper practical % limit to catching validly spoken
words, and the lower practical limit to the number of false positives
I'll see on other noises.

I see a lot of information about % recognition accuracy on the vendor
websites, but they refer mostly to noise-free environments and of
course to large dictionaries.

 

Maybe something here will help:
http://marf.sourceforge.net/

Good luck.
Richard

 

Lots to read here, thanks for the pointer. Not sure this will directly
give me the statistic I need, but I may be able to gather enough
samples of my "magic word" being spoken in different conditions to run
it through this s/w and generate some of my own stats.

 

Sounds a lot like "word spotting" of the old ( cold war ) days.
Lots of unencrypted voice radio transmissions in russian that
were recorded. People trained to listen in, identify in all
the uninteresting routine conversations key words.
Then that recording that possibly contained something
of value was further listened to by people who would actually
understand the language.
There was funding in the 70ies/80ies to do that inital
step cheaper by computer. Doubtfull if anything usefull
came out of it.

MfG JRD

 

It's not the same kind of application at all - really it's more like a
voice-operated "clapper" switch than anything else - but the
requirements are similar. The cost of a false negative or a false
positive are both pretty high, though a false negative is much more
costly.

I think cheap DSP technology has come a long way in the past 20-30
years

 

I'll take the muskrat Alex. Actually, it sounds like an old Firesign
Theatre line.

Scott

 

Never really got into Firesign Theatre. I prefer the Goon Show,
Hancock's Half Hour, etc.

Anyway, I was trying to demonstrate (flippantly) the real fact that
both a false hit and a false miss have real costs in this application
- a false miss is dangerous, a false hit is financially expensive.

 

Looks like Sphinx:

http://cmusphinx.sourceforge.net/html/cmusphinx.php

Funny enough, I heard of it first in the context of a speech/VR interface
for Infocom 'adventure'-type games.

 

Op Thu, 08 Jan 2009 17:48:34 +0100 schreef larwe <>:
I've got a very specific speech recognition application in mind, and
I'm looking for a reference that will indicate if it's feasible. I
want to recognize just one magic word, [...]

- If the computer doesn't recognize that you want its attention, a
ninja will beat you to death with a frozen muskrat, and
- If the computer hears your dog barking and thinks it was you trying
to get its attention, you'll be charged $1,000 for the CPU time.

Sounds like it could be a military application: firing too late is
dangerous and firing for no reason is costly. Or remote assistance:
screaming "help" too late is dangerous and rescueing you with a helicopter
for no reason is costly.

[...]

I see a lot of information about % recognition accuracy on the vendor
websites, but they refer mostly to noise-free environments and of
course to large dictionaries.

I think accuracy will be a lot higher in the case of whistled languages
like Silbo.
http://en.wikipedia.org/wiki/Silbo_Gomero_language

Can your users be trained to whistle?