Georgina Joyce wrote this on Sat, May 27, 2006 at 11:18:52PM +0100.
My reply is below.

> I wondered what the users of the system think about the voice?

Well, I can tell you that I think the *festival* documentation is
pretty dense.  From reading it, I'm not sure what the developers were
trying to accomplish.  I suppose it was English text-to-speech (TTS)
synthesis within reasonable storage and processing limits.  I suppose
they succeeded, sort of, for their day.  Things have moved on a bit
since then, and I would hope that better results were possible now,
but this surely is not my field, and I don't begin to understand its
complexity.

... so, if you take plain text in, you do well to put correct
pronunciation out, leaving aside how to pace the delivery, where to
place the emphasis, and how to vary the inflection.  Professional
actors are all over the map with their reading of Shakespeare, for
example.  You have to have some deeper concept of its meaning to
convey your sense of the text.  This, at this stage of the game, a
little computer cannot have.

... so, if you aim a bit lower just to automate the pronunciation, you
still need to simulate sense by varying the pace, emphasis, and
inflection while stringing together a sequence of phonemes.  It's
marvelous that it works at all, but the phonemes you choose to use
heavily are those that normally go together.  It bothers me that the
voices that *festival* provides tend to swallow trailing unaccented
syllables like -ble and -ing.  I suppose they do this so that such
syllables flow a little more naturally when they are not trailing.
Also, articles like "a" and "the" are given short shrift.  I suppose
this is because the emphasis given an article conveys a lot of the
sense in normal speech, and it's better not to try to convey anything
when the sense is unknown.  

... so *festival* generates a lot of pronunciation from plain text, and
it's intelligible -- barely.  It's like listening to a song in a
language a non-speaking singer has learned by rote: The cadence and
pitch are musically plausible but the sense of the lyrics doesn't come
through.  The listener has to take dictation in his mind's eye and
reread the text there to conjure the sense of it.

Somewhere along the line, I played with a TTS program that not only
read plain text but supported a phonetic input mode as well.  I can't
find support for any such thing in the turbid *festival* docs.  I
think it would be nice to spell the articles (thee, thuh, ae, uh) to
override *festival's* default pronunciation.  Surprisingly, this does
appear to work, but it's still too bad that *festival* doesn't provide
a list of acceptable spellings for phonemes.

-- 
.. Chuck Rhode, Sheboygan, WI, USA
.. 61F. Wind SSE 3 mph. Clear. Mist. 
-- 
http://linuxfromscratch.org/mailman/listinfo/blfs-support
FAQ: http://www.linuxfromscratch.org/blfs/faq.html
Unsubscribe: See the above information page

Reply via email to