Open desis123 opened 1 year ago
you probably have done something wrong,if you check here : https://www.kaggle.com/datasets/mobassir/comprehensive-bangla-tts you will see my trained log saved as .txt file and if you check my trainig log you won't see characters getting discarded, i have used this script to train vits with phoneme : https://github.com/mobassir94/comprehensive-bangla-tts/blob/main/bn_vits_tts/Bangla_phoneme_ViTS_trainer.ipynb if you are using exactly my dataset and my training script mentioned above then you shouldn't get character discarding warning,,make sure you follow the steps exactly like i did,,, i don't know any way of reducing training time.
Yes I have used exactly your dataset downloaded from kaggle .. Because I like to do it on your dataset first. Yes I might miss something. Let me try it from the beginning once again and yes I am using https://github.com/mobassir94/comprehensive-bangla-tts/blob/main/bn_vits_tts/Bangla_phoneme_ViTS_trainer.ipynb . Just changed the location parameter to run it . Still I might miss somewhere let me try it from the beginning once again . Thanks again for continuous help and support .
Make sure this code prints absolute path of espeak and espeak-ng : from shutil import which print(which('espeak')) print(which('espeak-ng'))
At least you should see path of espeak-ng Otherwise install espeak-ng by following Their documentation, then try to re run the code
Yes espeak-ng having absolute path , and after rerun the code having the same issue . But as I can see #phonemizer="espeak-ng",# multi_phonemizer espeak-ng is commented out here .. if that is the case is there any other uses of espeak-ng in other module ?
and aslo characters_config is commented out .. do you think that could cause this error ?
characters_config is commented out because we only need it while training without phoneme. the code you see in my training notebook is the exact same code that i used for training the model without facing any issue, you probably have issue with espeak installation or issue related to any package installation or version difference of coqui, this is the tutorial i followed for installing espeak-ng https://github.com/espeak-ng/espeak-ng/blob/master/docs/guide.md#linux
Thanks for reply .. I have just used following characters_config and sample down to 1000 just to see the quick result and affter just 500 epoch I have started to get some good result . With just 1000 sample epoch is taking like 30 seoncds right now . So it's just a result of 4-5 hours of training .. I am still continuing the training just to see optimal outcome of 1000 samples.
affer 500 epoch https://drive.google.com/file/d/1YsrikAe4vsAYCDO6eb-HQoN9L4ErRHz4/view?usp=drive_link
and character_config I have used
characters_config = CharactersConfig(
characters_class="TTS.tts.models.vits.VitsCharacters",
pad="
so far so good .. let see what came out after 5000 epoch .. I will keep updating here and if I have any question in future. and Lastly thanks for all help.
@desis123 good work. from my past experiments i have witnessed that with or without phoneme the performances are nearly same ,only difference is "you need to train longer for the model to converge if you do not use phoneme,but if you use phoneme then the model converges very fast(400-500 epoch is more than enough for phoneme model training)".
So you think it will not improve anymore ?? or should I continue at least 1000 ?
@desis123 as you are not using phoneme currently,so you can train more and hopefully performance will be improved,,,i said "if you use phoneme then you don't have to train for many epochs"
First of all thanks for getting a prompt reply for my previous question . I am sorry for disturbing you over and over . As I am pretty new to machine learning along with TTS . While I am running your training script at the beginning I am getting this following message
is it normal ?, I am not clear why characters not found is happening ?
second question after that script is running and currently on 256 epoch and each epoch is taking around 2 mins . I am running with 3090TI gpu , is there anyway to improve this timing ?