http://www.lojban.org/tiki/tiki-view_forum_thread.php?forumId=5&comments_parentId=6701
was the initial problem report.
I made a few changes and tested them and got feedback on it, with the
help of Chris Done hosting a few files at
http://www.jbotcan.org/jbo_test/ .
I received lots of useful feedback from that test, and I looked more
into espeaks c++ internals. I want to thank the people on the mailing
list and on irc for their assistance.
Some of the problems from that test was a general consensus that it
was too fast, and had other prosody issues like not honoring dot
pauses enough. Some thought it sounded like weird techno.
The "z" was usually not able to be heard correctly and many disliked
the "x" sound.
The lerfu pairs cyfy, syky, zydy, etc. sound all alike.
At least person also thought that my test text needed more sensible
sentences to test with.
I thought part of the reason it sounded bad was because espeak didn't
know word boundaries based on word morphology. That's where most of my
efforts since the time 2008-Aug-28 have been spent. I created a
preprocessing program named lihertadji.pl that splits words up and
puts in dots everywhere they are required. It should correctly handle
the test file jbo_test.txt .
However it is not complete and it has some hairy regexs that should
probably be double checked for correctness. `./lihertadji.pl <
jbo_test.txt > jbo_test0.txt` . I will include the file jbo_test0.txt
just in case someone doesn't want to run the script but still wants to
see it's output . I also made some more minor changes to jbo_list and
possible jbo_rules .
Splitting the words up should slow it down some plus I am trying to
give espeak an argument of 115 words per minute so it doesn't use
default 170 .
speak -f jbo_test0.txt -v jbo -w jbo_test.wav -s 115
that created a file with size around 32 GiB and 12 minutes long. I
won't ship it or any ogg speex files at this time because of size
issues and I have few more problems to solve before it again ready for
general testing. Maybe Chris Done will again host some files, I don't
know.
It might still have bad rhythm or prosody issues that will need more
testing and feedback.
I did nothing to fix 'z' or 'x' as of yet. I'm not exactly sure how to
proceed. z can be "z", "Z", "z." , "z;" or "Z;" afaik . I don't which
is z [z] a voiced alveolar sibilant . phoneme z claims to be
a "vcd alv frc sibilant" , I don't think the others match as well just
based on the label. Maybe it's now frictive when it shouldn't be??
``zoo'', ``hazard'', or ``fizz' . [½] is allowed z variant.
x [x] an unvoiced velar fricative -- not sure how to deal
with that at all.
phoneme x claims to be "vls vel frc" which is right on the money.
Maybe it just sounds odd to Americans. Maybe some of the other details
are wrong about it. I don't know.
phonemetable eo base
include ph_esperanto
phonemetable jbo eo
include ph_lojban
I think it uses esperanto to change the vowel sounds.
http://espeak.sourceforge.net/phontab.html mentions "phsource" and
"phonemes" files which is in espeakedit-1.38.zip I assume.
{cyfy} and friends get expanded to {. cy fy .} that might make them
clearer to hear. That being the case I added {la cyfyl e la sykyl e la
zydyl e la fybyzim e la cyzizam e la kygybyr cu jecta girzu} to the
test file to make sure I continue to give espeak a headache there.
However to my listening it still isn't always very clear, I think it
says the "by" lefru forms way too quickly.
As for sensible sentences I added lojban anthem into the test. If
anyone has some other suggestions as to what to add please be very
specific. I might also simply make another test file . This one has a
lot of synthetic word lists that tries to cover most of the sounds. It
aims for decent coverage more than being a realistic sample for
typical lojban speech.
http://allalone.org/cizra/ is where I got all letters test sentence
and the beginnings of the per letter tests and initial pairs test. I
wrapped many of those in {u'u lo'u ... y y le'u na gendra} so that the
whole file might be gendra .
http://allalone.org/cizra/sofybakni_slow.spx {.o'i mu xagji sofybakni
cu zvati le purdi } has all the letters from the lojban language.
speex 73K , 6 seconds . Look out! Five hungry Soviet cows are in the
garden!
http://allalone.org/cizra/sofybakni.spx speex 45K , 4 seconds.
http://allalone.org/cizra/rec.py the word list
alfas to zulus I got from Complete Lojban Language chapter 17
http://www.lojban.org/tiki/tiki-download_wiki_attachment.php?attId=181&page=The+Lojban+Reference+Grammar
.
I added a few words just to test a few things.
I added {a'o so'i da xamgu ma'a ma'a} just as another test sentence
which I used like the soviet cows one to see how it handles having the
space squeezed out of it.
http://www.lojban.org/tiki/tiki-index.php?page=Lojban%20Anthem is just
a slightly more realistic sample. The other stuff is from
http://stephen_pollei.home.comcast.net/~stephen_pollei/lojban/cilre_valsi/jeftu01.txt
, I changed some of the stuff that was quoting English, because I
didn't want to test mixing languages just yet.
Foot-notes:
http://video.google.com/videoplay?docid=-4507856202235021576 dead parrot
http://www.youtube.com/watch?v=J_nKJW_KuK4 fanmo jimte episode 1
http://www.youtube.com/watch?v=ZvODnSl7V3w fanmo jimte episode 2
http://www.youtube.com/watch?v=jvFRWpbZcWA u'icai .i ti bisli
http://selkik.podbean.com/
http://jbobac.lojban.org/
http://jbotcan.org/bacru/la-.an.-bebna.html
http://www.lojban.org/tiki/tiki-index.php?page=Multimedia
http://espeak.sourceforge.net/
http://espeak.sourceforge.net/add_language.html
http://espeak.sourceforge.net/editor_if.html
http://espeak.sourceforge.net/analyse.html
Attachment:
jbo_dictsource_2008-09-04.tar.gz
Description: GNU Zip compressed data