Re: IPA Text-to-Speech

Sally Caves wrote:
> I don't think this could work terribly well.  For now, at least.  Of course,
> what do I know about "Cool Edit" and the real cutting edge stuff out there?
> :)  In my experience, most contemporary voice recognition programs operate
> on a level far less complex than your actual enunciation (a step above
> lip-reading, to use an analogy), and I don't think anybody could honestly
> say, at this stage of our technological expertise, that the text readers are
> accurate renditions of pronunciation, although they are intelligible.  A
> program would have to be far more complex than even our most complete IPA to
> get timber, nuance, pitch, inflection, the minute changes that occur when a
> sound is proximate to another sound, and so forth.  Think of the subtle
> distinctions between a French "r" and a German "r"; and how in German the
> gutteral "r" (at least for me) will be different depending on what vowel or
> consonant comes in front of it (I can do a reasonable uvular trill for back
> vowels, but that changes for front vowels, and it's almost impossible for me
> after any consonant except "k"; after "t" it sounds like a French uvular
> scrape).   It will always sound like a machine speaking, with that
> machine-sameness.   And as for writing what you speak, think of all the
> kinks that have to be ironed out.  If it's overly sensitive, you'll get an
> unreadable bunch of gobbledygook depending on the speech patterns of any
> individual.  It's a great idea to aim for, but I think it will take a lot of
> new work to produce useable results, and until then I'd prefer to learn
> German pronunciation from an actual German.  But I can see the technical
> appeal of such a project.
I think a text reader for X-SAMPA wouldn't be impossible, especially if
it's strictly phonetic. The real problems start when you try to go from
phonemes to sound--you need a lot of language-specific data to make it
work. Of course, the utility of a phonetic text-to-speech program would
be a little limited, since you'd have to do all the work of converting
from a phonemic representation to a phonetic one yourself.

Reply