Xref: utzoo comp.ai:5029 comp.ai.neural-nets:1081
Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!ucbvax!agate!shelby!neon!Sunburn.Stanford.EDU!heck
From: heck@Sunburn.Stanford.EDU (Stefan P. Heck)
Newsgroups: comp.ai,comp.ai.neural-nets
Subject: Re: Backpropagation applications
Summary: NetTalk Accuracy
Keywords: Neural Networks, Efficient Learning
Message-ID: <1989Nov9.160406.14658@Neon.Stanford.EDU>
Date: 9 Nov 89 16:04:06 GMT
References: <1690@cod.NOSC.MIL> <77404@linus.UUCP> <13659@orstcs.CS.ORST.EDU>
Sender: USENET News System <news@Neon.Stanford.EDU>
Reply-To: heck@Sunburn.Stanford.EDU (Stefan P. Heck)
Distribution: usa
Organization: Computer Science Department, Stanford University
Lines: 9

According to Rumelhart in his ANN/PDP class here, Nettalk was trained on a 
set of the 1000 most common words rather than a random set. This run took
overnight to learn. They later also did a second test using 10 000 words. 
I don't know for which run the accuracy figures are, but supposedly it got
87% right except on words which were irregular. The best competitor at the
time was about 89% accurate. Human capability was estimated at 96%.

Stefan
CSD