Problems in tokenizing LibriTTS

[jry-king]

Thanks for your reproduction of the VALL-E paper! When I tried to prepare LibriTTS data with prepare.sh I encountered this problem:

I'm only using 4 tar files (train-clean-100, train-clean-360, test-clean and dev-clean) out of 7 in LibriTTS. Could you give me some suggestions about what's going on? Thanks!

[lifeiteng]

datasets fix a23f7c9

[jry-king]

Looks like the command is fixed but the "words count mismatch" warning still exists. I only found that it might be caused by some inconsistency between text files and phonemes. Will it actually influence the result?

[lifeiteng]

I haven't figured out these warnings yet, I checked the results and there is nothing wrong.

[pandamq]

@jry-king Hi, i meet the same warning info, i assume the reason is espeak tokenizer frequently take many words into one phone, e.g. ' of the' to 'ʌvðə'.
Do you have some solution to fix it, or we can igonre the warning to train the model correctly? Thanks!

[lifeiteng]

@jry-king Hi, i meet the same warning info, i assume the reason is espeak tokenizer frequently take many words into one phone, e.g. ' of the' to 'ʌvðə'. Do you have some solution to fix it, or we can igonre the warning to train the model correctly? Thanks!

can you provide more cases?

we need dig into https://github.com/bootphon/phonemizer.

[pandamq]

@jry-king Hi, i meet the same warning info, i assume the reason is espeak tokenizer frequently take many words into one phone, e.g. ' of the' to 'ʌvðə'. Do you have some solution to fix it, or we can igonre the warning to train the model correctly? Thanks!

it's hard to count them all. i can give you some cases.
"on the floor" to "ɔnðə floːɹ"
"had been" to "hɐdbɪn"
"of a" to "əvə"
"would have" to "wʊdhɐv"
"there is" to "ðɛɹɪz"

[lifeiteng]

errors shoud be fixed

ɜː -> ɜ ː

[lifeiteng]

@jry-king @mingqilearning Thank you for report these issues. It would be great if you guys could fix these issues.

Training looks good now #37

[pandamq]

@jry-king @mingqilearning Thank you for report these issues. It would be great if you guys could fix these issues.

Training looks good now #37

i tried to fix it, but failed. training good means the warning doesn't influent the training and inference?
it would be great if you share the phomezier and espeak-ng version.

[lifeiteng]

espeak-ng/espeak-ng#1703