Philip Bennefall wrote:
Hi again,
I solved my previous problem, turns out that I ran into the same issue as some other people (speech-tools was mismatched with festvox). I got it working now though, and am attempting to make a very very small ehmm voice. I only have 15 prompts in my inventory. It worked fine to label them, but along the way I got messages like:
Accepted sentences are: 15 / 15
Accepted sentences are: 11 / 15
Accepted sentences are: 8 / 15
Accepted sentences are: 6 / 15
Accepted sentences are: 5 / 15
Why does it disregard most of my sentenses after a few iterations? Could this be the reason for the issue below, so that I simply have to record more prompts to solve it?
15 sentences is almost certainly too few to get a good clustergen voice
(and certainly a good ehmm labeling, to start off with).
You'll really need to have a few hundred phonetically balanced sentences
for either ehmm or clustergen to have a chance of reasonable models.
You may be able to get a unit selection (for limit domain synthesis)
with 15 utterances, depending on the complexity of the utterances and
the types of things you want to synthesize. Following the talking clock
example (might) work.
If 15 sentences is all you have and you cannot get more, you could try
the voice conversion techniques (see festvox/src/vc/HOWTO and follow the
FESTVOX Transformation Voice instructions but with your recordings, but
again you really want phonetically balanced sentences to make theis work
reasonably.
Alan
I figured that since there were still some sentences left, I'd keep going, so I
ran the following seemingly without errors:
./bin/do_build build_utts
./bin/do_clustergen f0
./bin/do_clustergen mcep
./bin/do_clustergen voicing
./bin/do_clustergen combine_coeffs_v
./bin/do_clustergen generate_statenames
Though of course, it prints lots of information during certain stages so it is
possible that I have missed some errors, though I do not believe I did.
However, in the next step, it fails:
./bin/do_clustergen cluster
I get lots of errors like:
Clustergen f0 tree build on: ih_3
$ESTDIR/bin/wagon -ignore '(R:mcep_link.parent.lisp_duration)' -desc festival/clunits/f0.desc -data 'festival/feats/ih_3_f0.feats' -test 'fest
ival/feats/ih_3_f0.feats' -balance 0 -stop 200 -output 'festival/trees/ih_3_f0.tree'
/usr/tts/speech_tools/bin/wagon: input file not accessible "festival/feats/ih_3_f0.feats"
Type -h for help on options.
Clustergen f0 tree build on: ng_1
$ESTDIR/bin/wagon -ignore '(R:mcep_link.parent.lisp_duration)' -desc festival/clunits/f0.desc -data 'festival/feats/ng_1_f0.feats' -test 'fest
ival/feats/ng_1_f0.feats' -balance 0 -stop 200 -output 'festival/trees/ng_1_f0.tree'
/usr/tts/speech_tools/bin/wagon: input file not accessible "festival/feats/ng_1_f0.feats"
There is a lot more, but I had to trim it because my prior message didn't seem to get through.
I am not sure how to proceed at this point, so would be very greatful for some
tips.
Thanks!
Kind regards,
Philip Bennefall
_______________________________________________
Festlang-talk mailing list
[email protected]
https://lists.berlios.de/mailman/listinfo/festlang-talk
