Newsgroups: comp.speech
Path: lyra.csx.cam.ac.uk!doc.ic.ac.uk!agate!howland.reston.ans.net!cs.utexas.edu!chpc.utexas.edu!news.utdallas.edu!corpgate!nrtphaa9.nt.com!brtph560!brtph8ee!hoequist
From: hoequist@brtph8ee.bnr.ca (Charles Hoequist P250)
Subject: Re: Off the shelf Voice Recognition?
Message-ID: <1994Aug26.180853.7410@brtph560.bnr.ca>
Sender: hoequist@brtph8ee (Charles Hoequist P250)
Date: Fri, 26 Aug 1994 18:08:53 GMT
References:  <gbalfanzCv5A74.HI4@netcom.com>
Organization: BNR Inc., RTP, NC
Lines: 71

In article <gbalfanzCv5A74.HI4@netcom.com>, gbalfanz@netcom.com (Greg Balfanz) writes:
|> Does anyone have any feedback on off-the-shelf voice recognition systems?
|> 
|> 
|> I'm trying to get voice input for a UNIX database application.
|> 
|> I've seen Scott Instruments/VCS product at a trade show.
|> 	
|> 	alleged speaker independant, continuous?

I haven't played with their boards since the merger, and that
was before they were offering continuous. I tested them on a
bunch of 'yes' and 'no' tokens (telephone quality) and found
Scott could be tweaked to make it reasonably good.
Since my experience with their hardware is so old, I don't have
anything to say about current number of channels.
|> 
|> I've seen Creative Labs Voice Assist for sound blaster cards at a
|> computer store.
|> 
|> 	speaker dependant?
|> 
|> Any experience/feedback/alternatives?
|> 
Fun to play with. This is a stripped-down VPC image on the
SoundBlaster board (Creative Labs, the Sound Blaster folks,
have an equity stake in VPC).

|> 
|> Ideally, I'd like to NOT require a PC for every voice channel as
|> I'll have between 2-16 channels.
|> 
|> 
|> Thanks in advance!
|> 
|> Greg 
|> 
|> PS: My app will require RF (wireless) headsets for user interface.

I'd recommend checking into Voice Processing Corporation (VPC).
Their VPro system does speaker-independent recognition of yes/no
and continuous digits out of the box. I tested it on TIMIT digit
strings and found rates above 90%, but fluctuating some depending
on dialect and dropping drastically for children's voices. Also
tested it on yes/no tokens from telephone lines and found 96-98%
hits, weaker on 'no' than 'yes' and slightly weaker on women than men.
My group also tested it on digit strings over the phone and 
got good results. Note: you can improve overall string performance
a lot if you can put constraints on the digit string (e.g. has to be
a seven-digit string, or includes a checksum).

Regarding channels: VPC has several products which can either directly 
hook up to input (mike or RJ-11 jack) or can take multiple simultaneous
channels from other vendors' boards (Dialogic, Rhetorex, Natural
Microsystems).

My tech numbers for VPC are all obsolete, but you may be able
to get glossies and other marketing hype from 
Deborah Persson at 617-494-0100.

Hope this helps.

-- 
# # # # # # # # # # # # # # # # # # # # # # # # # #
Charles Hoequist, Jr.   | Internet: hoequist@bnr.ca
BNR, Inc.               |  voice: 919-991-8642
PO Box 13478            |  fax:   919-991-8008
Research Triangle Park NC 27709-3478
USA
The number you have dialed is imaginary. Please rotate your
telephone ninety degrees and try again.
