NetNews Usenet Archive 1993 #3/NN_1993_3.iso/spool/comp/ai/neuraln/4967

Text File | 1993-01-24 | 1.7 KB | 31 lines

Xref: sparky comp.ai.neural-nets:4967 comp.speech:489
Newsgroups: comp.ai.neural-nets,comp.speech
Path: sparky!uunet!gatech!darwin.sura.net!sgiblab!munnari.oz.au!bunyip.cc.uq.oz.au!cheops!pclink
From: [e-mail] (Rick)
Subject: Q: Teaching a neural net to speak with an accent.
Message-ID: <[e-mail]>
Organization: Telecom Australia, TNE Computer Support Services
Date: Sun, 24 Jan 1993 23:34:59 GMT
Lines: 20

I want my PC controlled answering machine to answer the phone with
the voice of Inigo Montoya (Princess Bride, wanders around saying
"'Ello.  My name is Inigo Montoya.  You killed my father.  Prepare to
die").  The problem is that during the entire movie, Inigo does not once
say "If you do not leave a message at the tone, prepare to die".  Last week
there was a show on the vidiot box that showed a neural net learning to
speak a phrase by feeding back the differnces between its attempt and the
utterance by a human.  My question is, is it possible to teach a net to the
point where it can correctly pronounce words it has not seen/heard?  A
related question arises because of the fact that the printed word has a low
information content compared to speech.  Is there a standard way of
encoding speech in text such that things like pitch, volume and emotional
content can be transmitted?  Is the above project feasible, or will the
computational requirements mean waiting until we have a Cray-in-a-wristwatch?
Note that I don't require real-time text to speech - I don't mind if the net
requires several 100M and a week to generate the speech, as long as I can
collect the output an play it back in real-time.

Thanks,
Rick.