Newsgroups: comp.publish.cdrom.multimedia,comp.publish.cdrom.software,comp.speech Path: cdrom.com!barrnet.net!decwrl!pa.dec.com!crl.dec.com!crl.dec.com!caen!uwm.edu!cs.utexas.edu!uunet!harlequin.com!dp From: dp@harlequin.com (Jeff Del Papa) Subject: Re: Speech Recognition Technologies In-Reply-To: ODonovan@McMaster.ca's message of Thu, 6 Jan 1994 19:35:27 GMT Message-ID: Lines: 43 Sender: usenet@harlequin.com (Usenet Maintainer) Nntp-Posting-Host: epcot.harlequin.com Organization: Harlequin Inc, Cambridge, MA References: <1994Jan3.180304.5618@netcon.smc.edu> <1994Jan4.064518.14953@microme.uucp> <1994Jan6.193527.12459@mcshub.dcss.mcmaster.ca> Date: Thu, 6 Jan 1994 16:02:00 GMT Xref: cdrom.com comp.publish.cdrom.multimedia:141 comp.publish.cdrom.software:133 comp.speech:192 This isn't on topic, but... >>>>> On Thu, 6 Jan 1994 19:35:27 GMT, ODonovan@McMaster.ca (Chris O'Donovan) said: Chris> Nntp-Posting-Host: mac-odonovan.physics.mcmaster.ca Chris> In article <1994Jan4.064518.14953@microme.uucp> Chris> jraven@microme.uucp (John Raven) writes: > It takes an extra card, runs on an IBM RS/6000 and costs just > under five thou. I think that is pretty much what it takes to > get good SR. Chris> For about $3000 you can get a complete Mac660AV which includes the Chris> computer (including CDROM, hard drive and 8Meg memory) _and_ the speech Chris> recognition hardware/software - the RS/6000 will run you another $10000 Chris> on top of the $5000 for the speech hardware (and everyone knows that Chris> while IBM hardware is great (read: fast) their software is horrible :) Chris> I won't evangelize, but you should check it out before wasting $15K on Chris> IBM hardware only to find that it only recognizes speech with an Chris> upstate NY accent. :) Actually the state of the art for big vocabularies is the stuff from dragon and kurzweil. The mac AV software is not sufficient for general dictation, as the vocabulary is much too small. To give it credit, it isn't intended for such use, it is supposed to be another way to enter commands. (the system I use, has an active vocabulary of 30,000 words from a 100,000 word dictionary.). The IBM system is interesting, and will be available for conventional hardware soon. I look forward to trying it. (I am typing injured, and am no longer allowed to use a keyboard. This message was entered with my hands sitting still in my lap.) Actually the hard part about developing speech software is acquiring a large enough sample of different accents to build your speech models from. With a large vocabulary, the last thing you want to do is train every word. (you can train at the rate of 6,000 words/day). It does put up a real barrier to entry for such companies, and I will hazard a guess that the libraries of voices are more valuable than the various speech recognition companies source code.