/proj/tts/data/turkish/babel/data/questions To be compatible with HTS, we had to rename some of the phones (e.g. '2') to spelled out versions (e.g. 'two'), since HTS expects alphabetic-only phoneme names. Not sure whether this is also required for Merlin.
Mert did a lot of debugging and data cleanup to get a Merlin baseline using the Babel conversational data. A modification to the Merlin code was required, to force it to output a model even when it would rather not (because the model is not so good) - see /local/users/mert/merlin/src/run_merlin.py on hecate, the commented part in line 331, for the fix - it forces Merlin to generate models regardless of validation errors.
A first baseline for Turkish Babel conversational is on hecate
under
/local/users/mert/merlin/egs/build_your_own_voice/s1/experiments/turkish_new
The voice was judged to not sound so good so some cleanup of the data was done. Mert's cleaned-up list of utterance transcriptions can be found here: /proj/tts/tools/babel_scripts/turkish_merlin/etc/txt.done.data
This consists of hand-corrected transcripts, and bad / unintelligible
utterances were removed. The beginnings of a voice trained on these is here, on
hecate: /local/users/mert/merlin/egs/build_your_own_voice/s1/experiments/turkish2
The label files and audio have a frame mismatch, and so need to be
regenerated to be able to build a voice.
I think we were using a "nonstandard" phone mapping for this data originally, so make sure to check the label files to make sure they are using ok phone names (i.e., matches other data you are using if combining). E.g., we were doing "oneone" instead of "oneCL".
Kai-Zhan created a Merlin baseline voice using this data; that voice
is here:
/proj/tts/habanero_archive/kl2792/tarballs/turkish_news.tgz
Erica has created a cleaned-up set of female utterances since the data
contains a lot of background music and lower-quality audio. That 4.5
hour set is defined here:
/proj/tts/data/turkish/bn/4.5hr_cleaned_f.scp
We also had to do the following phone mapping: