[slightly off-topic] programmable speech recognition software

[Fernando Pereira]

In article <3810bf00.16850204 at news.btx.dtag.de>, Stefan Franke
<spamfranke at bigfoot.de> wrote:

> On Fri, 22 Oct 1999 01:51:21 -0400, "Tim Peters"
> <tim_one at email.msn.com> wrote:
> 
> >What is "a room installation"?
> >
> Hmm, propably a messy homegrown German translation. Should mean a
> situation at a public exhibition. 
> 
> Our intention is to create a kind of interaction by listening to
> people with arranging microphones and processing the audio stream
> using several methods among which are pitch detection, speech
> recognition and recording/transformation/playback of sampled phrases.
This kind of environment is just about the worst possible for current
speech recognition technology. The problems include variation in
speaker position relative to microphones, multipath effects, noise,
background speech, variation in room acoustic properties as people move
about, just to mention a few problems.
> As you may anticipate from the above, it's not necessary for my
> purposes to recognize contiguous text. Some isolated keyphrases would
> be more than sufficient, even - say - special kinds of systematic
> errors or misinterpretations. My expectations are not as high as my
> initial posting may lead to suggest.
As far as I know, the best way to recognize anything in continuous
speech is to try to recognize everything, since the poor speech models
for each unit (phone, word) need all the help they can get in
constraining the places where they may match, and the best help we can
give them are the candidate matches of other models around them.

-- F