Build TTS-Thai to Flite voices.

[wannaphong]

I can't build flite voices. :( google/language-resources#31

[Shallowmallow]

Hi !
I'm actually maintaining an android version of flite on google play.
And I try to find all the festival languages ( and do so some too) and convert them so it can be accessible to everyone.
If you want I can adapt your voice to flite and tell you how I did it , once done.
But I didn't find the wavs on github

[wannaphong]

@Shallowmallow wavs file. (sample rate 16000 ) https://drive.google.com/file/d/1xI1TY54KLQKTsbPXmcEW0vyYeTWILuH0/view?usp=sharing

You can download the original file (TSynC-1) from www.nectec.or.th/corpus/index.php?league=sa. (CC BY-NC-SA 3.0)

OUTPUT_VOICE_FOLDER : https://drive.google.com/file/d/1i5z8yjXKBDzfc3zbaJJhg9L2hE5f4byC/view?usp=sharing

Rule : https://github.com/PyThaiNLP/tts-thai/tree/build2/th

[Shallowmallow]

Great. Thanks was just able to download it.

[Shallowmallow]

Oh I looked at your files and saw you were only using a lexicon and not some rules for unknown words. Are there easy to rules to convert an unknown word to their phonemes ? :)

[wannaphong]

@Shallowmallow You may have to convert unknown words to IPA first. I have a problem not being able to convert such unknown words as well. Thai language, we don't have enough public dataset to do G2P with a deep learning.

[wannaphong]

@Shallowmallow cmu_th.flitevox not working. :(

Segmentation fault (core dumped)

• cmu_th.flitevox : https://github.com/wannaphong/flite/blob/master/cmu_th.flitevox
• cmu_th_lang : https://github.com/wannaphong/flite/tree/master/lang/cmu_th_lang
• cmu_th_lex : https://github.com/wannaphong/flite/tree/master/lang/cmu_th_lex

[Shallowmallow]

Hi !
@wannaphong
So the words only came from a dictionary ? Is thai easy to read ? Do you think the rules are simple to make , without using any deep learning.

I'm trying two things :
- doing the thai with festival grapheme scripts. It will resolve the rules problem.
For now, it works well when I use 200, 500 sentences but when I do all 2000 thousand sentences, there are horrible parasite sounds. So I'm testing 1000 , then more.
- converting the actual voice to flite. I can convert easily the voice to flite but it talks gibberish as I have some problems with converting the dictionary. ( the first time I'm doing this with pure unicode :).
But I actually think , there is another problem, I haven't checked throroughly yet. THere may be a problem because of the phonemes name ( specifically those ending with numbers). Flite usually converts stress to numbers. So for example (( b e t) 1) will become b e1 t .

PS : I remember you saying you had a voice that wasn't using unicode. Do you still have the txt.done.data for this voice ?

I wanted to test the festival grapheme conversion

[wannaphong]

Hi !
@wannaphong
So the words only came from a dictionary ? Is thai easy to read ? Do you think the rules are simple to make , without using any deep learning.

I'm trying two things :

• doing the thai with festival grapheme scripts. It will resolve the rules problem.
  For now, it works well when I use 200, 500 sentences but when I do all 2000 thousand sentences, there are horrible parasite sounds. So I'm testing 1000 , then more.
• converting the actual voice to flite. I can convert easily the voice to flite but it talks gibberish as I have some problems with converting the dictionary. ( the first time I'm doing this with pure unicode :).
  But I actually think , there is another problem, I haven't checked throroughly yet. THere may be a problem because of the phonemes name ( specifically those ending with numbers). Flite usually converts stress to numbers. So for example (( b e t) 1) will become b e1 t .

PS : I remember you saying you had a voice that wasn't using unicode. Do you still have the txt.done.data for this voice ?

I wanted to test the festival grapheme conversion

txt.done.data file in thaitts2.zip. https://drive.google.com/file/d/1i5z8yjXKBDzfc3zbaJJhg9L2hE5f4byC/view?usp=sharing

[Shallowmallow]

I tried making a festival grapheme voicee http://www.festvox.org/festvox/c3485.html .
Because like this there would be some rules.
It uses this converter https://github.com/nltk/nltk_contrib/blob/master/nltk_contrib/scripttranscriber/Unitran/Tables/Thai_unicode.txt
For some reason , the clustergen didn't work well ( there were some noises times to times when I did more than 500 sentences), so I tried to do a HTS voice.
But the sentences are not really correct ( I think).
Festival_grapheme is used here http://festvox.org/cmu_wilderness/
But the thai doesn't seem correct . It has a distortion measure MCD of 6.98 . Can you check if you understand the sentences on the website ?

I think the problem with the grapheme voice and with your dictionary is that you consider a tone like a phoneme.

It appears okay with your voice , but only because the words appear exactly as they have been learned.

But logically, it should be a problem.
Because the sound will be cut into phonemes.

So the sound for example à which is a 4

will be cut like this
|a|__4|

instead of
|a4_______]

so basically if you prononce ì it will do
|i|__4|
except that 4 sounds like the end of a so it will be kind of ià.

I see 2 possibilities to correct it

• don't consider a tone a phoneme
• consider it a phoneme but force it's duration to 0 . Hmm I wonder if you can do it easily with festival.

Do you plan to use the voice even if doesn't rules ro read unknown word ?

[wannaphong]

@Shallowmallow I was thinking of use g2p and use SayPhone on festival.