Noob need help with Mozilla TTS

TTS (Text-to-Speech)
#1

First of all, i am a total noob, so very sorry about the basic questions.

I am training a custom dataset based on the LJspeech structure using Tacotron2.
I am at 170k iterations after using a pretrained LJspeech model.

I get the preview audio for each epoch, but my question is now; how do i use it with other vocoders than the Griffin Lim one. I know this is basic for most of you but i havent coded a single row of python before starting with this.

Are there any basic tutorials for this, or even the syntax to generate the voice from a text string using some of the other vocoders?

I am using Colab for training and have setup a notebook for the purpose.

#2

Hi! There are some notebooks in the repo. In there there are some pretrained vocoders you can try with your model. Just take the notebook and point it to your model when calling the TTS checkpoint instead of downloading the pretrained one. Mind you they have to have been trained using the same audio attributes.

#3

Hi, Thanks for the answer. Wouldnt i have to also download these pretrained models and upload to the notebook i am running, or will i get them when i clone the repo?

#4

So is a part of the workflow also to train the actual vocoder model (if the audio parameters do not match)?

How would i do that?

I use this command to train the tts model

!python /content/TTS/mozilla_voice_tts/bin/train_tts.py --restore_path “/xxx/checkpoint_170000.pth.tar” --config_path “/xxx/config.json”