Compaq.com - About SpeechBot

Ernest Perez eperez at sparkie.osl.state.or.us
Wed Dec 15 20:16:18 EST 1999


Web4Lib folk,

Stephen Sloan comments negatively on the ability of the Compaq SpeechBot approach to handle personal name recognition and the quality of transcription. Agreed. Especially since this is an early stage of development, not a finished mainstream product. I don't believe it was really put up as the latter. I see and concede his points. 

But just a moment...let's remember exactly what we're doing here? Like they say, "SpeechBot is an experimental index of popular U.S. radioshows, based on state-of-the-art speech recognition technology." 

Look, we're talking about an approach to searching audio text. We're _not_ talking about name entry control, or quality comparison to OCR, or the selectivity of their PBS test datafile. We're talking about the developing the kind of "best match" or "near miss" kind of recognition for spoken audio, that products like Excalibur have used in text searching. Like Roy Tennant suggested, look for ...
	Dr. Laura library libraries filter filtering
and you'll find her stuff. Yeah, I admit it _looks_ crummy, perhaps, but I want to be able to search it easily and quickly and find it and _listen_ to it, not print it out or get a quality text file! You go to a transcription service if that's what's on your mind!

Another example, try looking for ...
	breast implant litigation 		[all of the words]
And see whether you can quickly find some broadcast text that's relevant. I don't have any better alternative ways on me at the moment.

My mention of it to Web4Lib was meant as a pointer to an interesting software technology development area, and possibilities. I was not endorsing a transcription service or a name authority control utility.  Hope it was of some potential interest. For example, when I sent the same note to my friend the Oregon State Printer, his forwarding note to our colleagues at the State CIO's office, the Legislature, etc., was...

------------------------------------
Check out this link, forwarded from Ernest Perez. This search engine does
speech recognition and indexing of audio files. Lots of applications for
us: public meetings, legislative hearings, etc.
------------------------------------

Cheers,
- ernest
______________________
Ernest Perez, Ph.D.
Group Leader
Oregon State Library
503-378-4243, ext 257
ernest.r.perez at state.or.us

====================ORIGINAL MESSAGE===========================

On Tue, 14 Dec 1999, Ernest Perez wrote:

> <http://speechbot.research.compaq.com/cgi-bin/query?help=about>
> 
> Slick! Really slick! Free text search of audio text. It's a great beginning.
> 
A very small beginning, I think. There are several problems.

Speech recognition tools have problems with personal names. I once used
such software to make a listing of a poetry collection donated to our
library. It was very useful as I could handle each volume and record
bibliographic information as I examined the volumes. Recording authors
was a problem as the program always hazards a guess. I wish I could
remember Naturally Speaking's stab at "Siegfried Sassoon". All I can
recall is that it was pretty hilarious. This databse has similar
problems. I can't find any reference to John Olerud. He must have been
discussed in the sports show as he signed as a free agent recently. I
also tried searching "rude" and came up empty.

Aside from personal names, looking at the "transcripts" from these shows
illustrates how badly the recognition software can work. Some of the
material is just incomprehensible. Compaq acknowledges this in their
information files. They say that most important words are spoken more
than once and the software will get it right eventually. It would be
interesting to compare this approach to OCR. CIHM has scanned microfilm
and indexed the OCR'd results. Would a project work better if someone
read the text into speech recognition software? The CIHM project is at:

http://www.canadiana.org/

Here, the dirty OCR is hidden from the viewer. It's used for searching
only. The viewer sees page images or PDF files.

I'd also like to comment on Compaq'a decision to devote 10% of their
efforts on this project to shows about the paranormal. I won't, however,
as this would probably cause an off-topic flame war that would resolve
nothing. Let's just say I was surprised to see 2 shows listed here. I
guess that in the backwaters of New Brunswick we don't get to listen to
such .... er... programming ....much.

Stephen Sloan
Systems Librarian
UNB Libraries
sloan at unb.ca
(506) 453-4814



More information about the Web4lib mailing list