Date: Fri, 14 May 1993 03:14:35 GMT From: dp@world.std.com (Jeff DelPapa) Subject: Impressions: Kurzweil VOICE Organization: The World Public Access UNIX, Brookline, MA Newsgroups: sci.med.occupational I recently got to play with the Kursweil voice system. -- the MIT access office had deuling voice systems day, and as a current DragonDictate user, I was quite curious. Quick summary: I was impressed, I occaisionaly entertain thoughts of jumping ship. Not quite so quick summary: Random spec like things: taken from lit and sales manager comments, Cost $6000. Vocabulary 50,000 words, 10,000 user defined, 40,000 from a 200,000 word dictionary. discrete utterance, minimal training to start, (surprisingly low, claiming speaker independence may be stretching it, but I did watch the hordes try it with fair success), continusly adaptive voice models (at 12mb they are much larger than dragons). Some context sensitivity (tries limited language model to get some of the homophonic words right, adapts some of the commands depending on which program is running (you say save-as, it turns it into the correct keystokes for word or word perfect, or lotus depending on which it thinks is running, if it thinks you are running lotus, it always does digits rather than One, Two..)) MS-DOS only, tho appears to work with some protected mode programs. windows is "soon" -- While they did use DesqView for task switching, they didn't know about DV/X and the hands free X stuff -- since they are local, I intend to pay them a visit, with my current setup. Requires: 486DX 33mhz or better, 32mb memory. Has some interface stuff for "home control" (usefull for those with minimal arm mobility) One full length 8 bit ISA slot, board looked fairly heavily populated (I muttered PCMCMIA at them, but agreed that they would have their work cut out for them getting the power consumtion down). Available with a telephone headset or a boom mike. They actually occaosionally beep in your ear which I found annoying (presumably it can be shut off, with wire cutters if nothing else) The telephone handset had some problems with ambient noise (we were in an alcove just off the infinite corridor at MIT (the athena "fishbowl"), and the noise of class changing traffic gave it problems (it was loud enough to require raising ones voice to talk to humans), at a noise level that would have been a loud office (during class), it didn't seem to mind, and the boom mike (noise canceling like the dragon uses) did better in all situations) impressions: I watched a number try it, and tried it myself. It did surprisingly well with random adult american males that walked up to it. (better than dragon does untrained) It did fairly well with me, tho my using command phrases tended to throw it -- as a current voice recognition user, I tend to clip short the typical command phrase (anyone who has to work near a DragonDictate user will be quite familliar with "gotosleep" and "beginspellmode" uttered as if a burst from a machine gun. The "train" mode has a display of utterance length vs model, and on many words I was near the NA average, but on phrases, I was a minimum of 20% faster than the norm - I tried it with the salesmans model, and did better, as his pronunciation speed more closely matched mine) I still wonder about the company stability (dragon has been around for a decade, Mr. Kurzweil has changed direction a few times), and it wasn't available when I chose Dragon (it was released in january), but I would really consider it if I were making the decision today (in fact if I wern't still paying for my system and the machine it calls home, I would be tempted to jump ship)-- if the vocabulary size is real, it would be helpful -- (I am over 1000 user defined words already, and that is because I am conscious to prune things - I couldn't fit the lisp reserved word list into either machines user defined vocabulary, so I havent tried) contact: Kurzweil Applied Intelligence, Inc 411 Waverley Oaks Road Waltham MA 02154 617 893 5151 fax 893 6525 quick spec comarison with DragonDictate (if not mentioned, I considered them similar enough)-- dd is $4995, requres a 386/20 and 8mb of memory (I haven't tried it on a machine slower than a 486/33), doesn't require a FPU. 30,000 word vocabulary, 5,000 for the user, 25,000 out of a 70,000 word dictionary. Moderate training (200 words). Doesn't like protected mode programs, windows has been "soon" for more than 1 year, program has been commercially available for at least 2 years. -- It is possible to have program specific vocabulary, tho you have to explicitly load and unload it. not a whole lot of context sensitivity (mostly just if you are entering digits, leaves out spaces, doesn't use spelled out word) works (tho I consider how it works a kludge) with DVX well enough that a2x with a unix box is at least as good as using the stuff on the DOS box natively. Usual disclaimer: I have no connection with either company, tho I do own a DragonDictate system. This may be re-printed without restriction.