Tts: Is there any interest in providing a model trained in Brazilian Portuguese?

[Edresson]

Hello, I recently developed a dataset for voice synthesis in Brazilian Portuguese using my own voice, called TTS-Porguese Corpus, the base has approximately 10 hours of talk, is available at: https://github.com/Edresson/TTS-Portuguese-Corpus

I have already successfully trained the DCTTS model in the dataset.

I created a branch in my repository to support the dataset in this repository, see here: https://github.com/Edresson/TTS/tree/TTS-Portuguese

In initial tests the model unfortunately did not converge, at the moment I'm out of time, and machine to train the model.

  Is anyone interested in training/adjusting the hyperparameters of the model to get a TTS model in Portuguese?

[erogol]

Unfortunately not at the moment. Maybe later on.

[Edresson]

Hello, I trained the model with the TTS -Portuguese Corpus. Provide a demo here: https://colab.research.google.com/drive/1OZ7T5fqR-QPyN5O6HFetkOeHNCKbdGsA

[erogol]

would you mind if I link your model on our released models https://github.com/mozilla/TTS/wiki/Released-Models

[erogol]

@Edresson have you tried universal wavernn trained here https://github.com/erogol/WaveRNN? It sounds already good to me but that might improve the results with neural vocoder.

[Edresson]

@erogol you can link my models.
I trained two models. The first model was trained only under the TTS-Portuguese corpus (demo link ). The second model was trained using transfer learning (with the model iter-120K | LJSpeech | commit: bf7590) demo link

About WaveRNN I had been trying to train at TTS-Portuguese Corpus, however, I didn't get good results (I wasn't using transfer learning ). It would be interesting to try its implementation.

[erogol]

@Edresson which version trained better, the one with transfer learning or the other?

[Edresson]

@erogol the version with transfer learning is better.

[Edresson]

I will try to use Universal WaveRNN.

[Edresson]

@erogol @reuben New pretreined model added to Colab Notebook ( link ). The notebook generates audio synthesized by the WaveRNN and Griffin-Lim vocoder. WaveRNN audios sound better.
Some issues in stop token loss have been fixed, now the model correctly predicts the pronunciation time of each sentence. I believe it will be the final model for the TTS-Portuguese Corpus. If I have new results I will add it here.

Below the details of the model if you want to add it on the released models page.

Audio Sample
Pretrained Model: Tacotron-iter-381000
Details Model: Tacotron + Universal WaveRNN for Brazilian Portuguese
Commit: Branch
Dataset: TTS-Portuguese

Note: The model was trained using transfer learning ( with the model iter-120K | LJSpeech | commit: bf7590 )