Theiling Online    Sitemap    Conlang Mailing List HQ   

Speech synthesis

From:Mark J. Reed <markjreed@...>
Date:Tuesday, May 2, 2006, 14:34
So I have one of those automotive GPS units that speaks directions.
Since the database of place names is all text, such directions involve
text-to-speech.  And I must say, I'm impressed at how natural it
sounds.  I mean, it's still obviously artificial, but we've come a
long way from SAM.

Things I've noticed:

1. It doesn't make subphonemic positional distinctions.  Initial /l/
is [L] instead of [l], and initial /t/ is [t] instead of [t_h].  The
latter is especially obvious in a phrase that occurs on my way to
work: "Turn right on Tabernacle St.".  As part of the core vocabulary,
|Turn| is stored phonetically, and therefore has aspiration on the
/t/, but |Tabernacle| does not.

However, it does make positional distinctions when the two phones
contrast phonemically elsewhere.  Thus, for instance, "Battle" is
correctly (for American) pronounced /badL=/ rather than */batL=/.

2. It generally does a good job guessing emphasis, but doesn't believe
in multiple primary stress.  Thus, "I-75 South" (that is, the
southbound lanes of Interstate highway number 75) is pronounced
"eye-seventy-FIVE-south", rather than the more natural-sounding "EYE
seventy-FIVE SOUTH".

3. There are some odd interference effects.  I'm not sure where these
come from.  I assume it makes some attempt to emulate assimilaton,
because the flow is pretty good, but there are some really weird
artifacts.  For instance, after a road name that ends in /l/, the /d/
in the word "road" comes out more like /z/ - e.g. "Roswell Road" comes
out sounding like "Roswell Rose".  Any ideas on what could cause that
would be welcome...

4. But by far the most amusing utterances result from incorrect
acronym expansion.  Due to a lack of imagination, several  minor roads
around here - roads whose primary purpose is to provide a connection
between two more major roads - have names like "Canton Road
Connector".  These are abbreviated "Canton Rd Conn" in the database,
which the unit cheerfully pronounces as "Canton Road Connecticut".  :)
Similarly, the various Ind.(ustrial) roads have "Indiana" in their
names...

Anyone here do research in this area?  It must be a fascinating field.
--
Mark J. Reed <markjreed@...>