TTS: [Bug] Tacotron 2 with DCA is broken

🐛 Description

I’ve trained vanilla T2+DCA for 300k steps on the current main branch and it sounds/looks broken.

To Reproduce

Training script Config

Environment

Environment

Additional context

Audio files from 300k steps

Screenshot 2022-02-21 at 10 52 15 Screenshot 2022-02-21 at 10 52 21 Screenshot 2022-02-21 at 10 52 08 Screenshot 2022-02-21 at 10 51 58

Train alignment figure:

trainfigure

About this issue

  • Original URL
  • State: closed
  • Created 2 years ago
  • Comments: 15 (13 by maintainers)

Most upvoted comments

all is good, closing this. Thanks @lexkoro

+1
a-froghyar on Feb 23, 2022

@erogol perfect thank you! I’m waiting for my training to run, posting results here tonight and closing the issue.

+1
a-froghyar on Feb 23, 2022

This recipe works flawlessly

https://github.com/coqui-ai/TTS/blob/UpdateDCARecipe/recipes/ljspeech/tacotron2-DCA/train_tacotron_dca.py Screenshot 2022-02-23 at 14 56 53 Screenshot 2022-02-23 at 14 57 18

+1
erogol on Feb 23, 2022

Might be because you are using np.log with spec_gain: 20, either try

spec_gain=20.0,
log_func="np.log10",

or

spec_gain=1.0,
log_func="np.log",
+1
lexkoro on Feb 22, 2022

Started to train the LJSpeech DCA recipe. I share my updates here if there is an issue or not

+1
erogol on Feb 22, 2022
Read more comments on GitHub
← TTS: [Bug] RuntimeError: CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling `cublasSgemm( handle, opa, opb, m, n, k, &alpha, a, lda, b, ldb, &beta, c, ldc)` when training tacotron2
TTS: [Bug] The model does not train with the new hyperparameters given by command line when trying to restart a training with `restore_path` and `continue_path`  →