[question] training chinese tacotron2 model with my own wavs

[hudsonx1123]

Hi,
I try to train chinese tacotron2 model with baker.
I replaced the wav files with my own wav files and re-define 000001-010000.txt following original format.
It's ok in preprocess and traing progress, but something is wrong when I copy the model to sample baker colab for test.
Input is about 20 chinese-words sentence.
It is just about 6 sec sentnece but Tacotron2 + MB-MelGAN generate 1:39 wav(as attachated image).
It can correctly speak the sentence in the first 6 sec but strange sound in the other duration.
any suggestion about this problem??
Thanks a lot!!

[azraelkuan]

@hudsonx1123 seems you train a bad tacotron, can u share me you Groud Truth Align mels or Train result using tacotron.
u can check it in your log dir

[hudsonx1123]

@hudsonx1123 seems you train a bad tacotron, can u share me you Groud Truth Align mels or Train result using tacotron.
u can check it in your log dir

@azraelkuan ,here are image and event data of my train result(I trained 80k steps this time). Thanks for your reply and help.

events.out.tfevents.1604042982.ND6TensorFlowAndCaldi.24689.505.v2.zip

[azraelkuan]

it seems that your have some bad data in you dataset, u need to filter it out, or add grad norm to handle this.

[ahmedalbahnasawy]

@hudsonx1123 It might be data issue. What is your batch_size and ngpus ?

[hudsonx1123]

it seems that your have some bad data in you dataset, u need to filter it out, or add grad norm to handle this.

@azraelkuan I tried to remove all the Chinese period symbol in the metadata file (000001-010000.txt) and train again.
for example:
000001 卡尔普#2陪外孙#1玩滑梯#4。
to
000001 卡尔普#2陪外孙#1玩滑梯#4

Then the issue was fixed and I get good result.
Thanks for your help!!

[hudsonx1123]

@hudsonx1123 It might be data issue. What is your batch_size and ngpus ?

@ahmedalbahnasawy I just use default baker conf(tacotron2.baker.v1.yaml) to do training.