[comp.cog-eng] Speech Recognition

reiter@babbage.harvard.edu (Ehud Reiter) (01/11/89)

In article <4629@xenna.Encore.COM> bzs@Encore.COM (Barry Shein) writes:
>Re: need a teraflop to do human speech recognition...
>Last time someone told me this I assured them that if they could
>cobble together something which would recognize human speech
>accurately on a more conventional machine but veerrry slowwly I would
>find them a few hundred million to merely speed it up, no problem.

More processing power does in fact improve performance on speech recognition
systems, at least the ones I happen to be familiar with.  If nothing else,
more crunching power lets the system do a more thorough search when it tries
to put phonemes together into words, or words together into sentences.

The question is not whether more processing power improves recognition
accuracy (it does), but whether it improves it *enough* to make the
system useful in a real application, *and* also whether people will be
willing to *pay* for the extra processing power.  No one will pay $100
million for a speech recognition system - they'll learn to type
instead, or maybe hire a secretary.

My impression is that the slow but steady progress of speech systems owes
as much to the steady shrinking of $/MegaFlop as it does to the improvement
of recognition algorithms.

					Ehud Reiter
					reiter@harvard	(ARPA,BITNET,UUCP)
					reiter@harvard.harvard.EDU  (new ARPA)