Train Glow TTS for Portuguese #500
Replies: 2 comments 1 reply
-
maybe @Edresson can help you as he is the most proficient person training Portuguese models using 🐸 TTS. |
Beta Was this translation helpful? Give feedback.
-
@erogol and @Edresson I am trying to train a Multispeaker Portuguese Model. So, I am finetuning (restore_path) the Portuguese Single speaker Model from this notebook (https://colab.research.google.com/drive/1sh96-f9J2BJcLm1rz5KBkwq3YP83bEr5?usp=sharing) and adapting the config for using in a multispeaker way. The speaker embeddings were generated from the mueller91's model. I am training using CommonVoice. The loss is decreasing, but the synthesized utterances appear to remain related with the single speaker initial model (the voice is similar to that single speaker, instead of variable across the speakers). :/ I am wondering if that is related with some error I am making in config. So I would like to ask for help. My questions are:
Thank you!! Reference used: Speaker Enc model (https://drive.google.com/drive/folders/15oeBYf6Qn1edONkVLXe82MzdIi3O_9m3) |
Beta Was this translation helpful? Give feedback.
-
I am interested in training Glow TTS / Tacotron2 DDC for Portuguese language.
I couldnt find proper instructions about the changings I have to do for it, especially how to choose the phonemes to use in the config file.
I would also to perform transfer learning if it is possible, as @Edresson has done in his paper (https://arxiv.org/pdf/2005.05144.pdf) to decrease the training time., but it is not a hard requirement for me.
May someone help me about what I have to do in order to train GlowTTS in Portuguese , please?
I am planning to use Common Voice. (I have also thought about MLS, but it has not punctuation)
Beta Was this translation helpful? Give feedback.
All reactions