This is One-Step... in view mode; [Up]
Date: Sun 04-Jan-1989 07:45:48 From: Unknown Subject: Re: One Step... (long!) In article <397@laic.UUCP> darin@laic.UUCP (Darin Johnson) writes: >>On the other hand....human like speech recognition has been postulated to >>take on the order of a Tera Flop by the IBM people. >This is if you use the standard Von-Neumann architecture, standard >algorithm's etc. I saw a setup awhile back at UCSD that did speach >recognition using neural networks (P.D.P. for you purists). Although I >never actually saw it run (only graphics output), it was supposed to be >able to 'decode' sentences in roughly 1/4 real time. Presumably, this >was with ideal conditions, short sentences, etc. Boy, do I have some grant proposals I'd like *you* to review. You've been duped on this one. Neural nets have been used to do speech recognition but I haven't seen a neural net that performs nearly as well as the Hidden Markov models that are now the state of the art. The neural net you saw probably accepted input symbols at a low rate (100Hz). Getting the data that the neural net needs as input (10ms of audio signal converted into a single vector in a small dimensional space) is a pretty hard problem. They probably used LPC and this completely ignores the issues of speaker adaptation, multiple speakers and some other very hard (unsolved) problems. > With a hardware neural >net, real time speach recognition is quite possible with less than a >super-computer. You wouldn't even have to devote an entire machine room >to it. Only if you count the neurons between your ears as a neural net. It is important to realize that 80% or so word recognition rates are relatively easy to attain but cutting your error rate by a factor of two gets progressively harder. This is the reason that speech recognition has been so tantalizingly close for so long. The amazing thing about neural nets is that people have got some good results without putting much knowlege (or understanding) into the model. Whether neural nets (aka ignorance engineering) will do better than more carefully crafted systems remains to be seen. Malcolm >From: daveh@cbmvax.UUCP (Dave Haynie)
These are the contents of the former NiCE NeXT User Group NeXTSTEP/OpenStep software archive, currently hosted by Marcel Waldvogel and Netfuture.ch.