🚀 go-pugleaf

In article <1990Mar9.144541.29446@spectre.ccsf.caltech.edu>
gbrown@tybalt.caltech.edu (Glenn C. Brown) writes:
>	The really depressing thing to me is that I hear (from a very reliable
>source) that someGuy@Apple ran some sort of neural net on a phonetic
>dictionary (using apple's CRAY) and manged to distill the pronunciation rules
>of English (even the irregular stuff) down to a file of < 100k.

So?  DECtalk (Dec's speech synthesizer product) has very good pronunciation
rules using a 64 kbyte "intelligent" dictionary and a fairly small program
to run it.  All of DECtalk runs on a 68000, with a separate signal processor
(TI 320) to do the speech generation.

From having listened to it, I suspect that Macintalk contains a simple
text-to-speech program (perhaps using the public-domain NRL rules that can
be implemented in one page of Basic) and a series of waveform tables that
are fed into the sound generator (the text to speech program builds a stream
of waveforms). If the company that wrote Macintalk didn't license/sell the
waveforms to Apple,  they might not be able to change them to suit later
hardware.

I suspect that a crude speech synthesizer could be brought up on one
of the Mac-II series without too much difficulty, however the quality
of DECtalk demands much more compute resources than you can get on a Mac
(to do the actual signal generation).  I'm suprised that NeXt hasn't
done one: perhaps decent quality is harder than you software folk
realize.

Martin Minow (ex DecTalk developer)
minow@thundr.enet.dec.com
The above does not represent the position of Digital Equipment Corporation