Raping mbrdico into a tts pipe.

[Hans Zoebelein <hzo goldfish cube net>]

Hi Blinuxers,

as you may know there is mbrdico which is thought by its father
Vincent Pagel (http://tcts.fpms.ac.be/~pagel) as 'a talking dictionary
using MBROLA'.

If you want to understand how the MBRDICO letter to phoneme algorithm
works, read "LETTER TO SOUND RULES FOR ACCENTED LEXICON COMPRESSION", by
Vincent Pagel, Kevin Lenzo and Alan W. Black Proc ICSLP98 Sydney.

Mbrdico is normally called once and you get an input line which is
spoken after you hit <return> through Mbrola.

But it seems that  mbrdico is also a very stable text to speech
processor. Much more reliable than approaches like 'freephone'.

Brutally piping ASCII output into mbrdico makes it a very reliable 
text to speech tool. You may test it by yourself. A fast hacked script
could look like this:

[begin script]
#!/bin/sh

cd /usr/src1/mbrdico/mbrdico/bin
export us2="mbrola  $* /usr/lib/mbrola/us2/us2 - - | rawplay "
./mbrdico.Linux american.ini 1>/dev/null 2>&1
[end script]

Calling this script as mbrdico.go, you simply can read a text by
'cat <your_text> | mbrdico.go'

Screader screen reader works fine with mbrdico. Other screen readers
should too. Mbrdico also reads even huge files without ever been killed
in between.

Adding some kind of server client approach and improving
handling of punctuation through a preprocessor seems to be easy.
And as a general text to speech server seems still far away...

Check it out. Mbrdico is available as
ftp://tcts.fpms.ac.be/pub/mbrola/TTS/MBRDICO/MBRDICO990625.zip
or at the Blinux file archive as
http://leb.net/blinux/mbrola/tcts.mirror/TTS/MBRDICO/MBRDICO990625.zip

PS: Mbrdico can be trained to digest new languages. It uses a neural
network approach. French, american and British English dictionaries are
available on the net to train the little beast (See
http://tcts.fpms.ac.be/synthesis/mbrdico/ 
for details


Enjoy! 

Hans 
(maintainer blinux-lists)