Flux9665
commented
1 year ago Hi! Yes, enabling training models for new languages is one of the core goals of this repo. I tried my best to write detailed instructions in the README on how to train a new model on a new language.
Generally you don't need to train an aligner and a vocoder, those two components can just be downloaded from the release page and used as they are. For the TTS model that goes from text to spectrogram frames, you need to train a model. You can either do that from scratch, or you can download the meta-learning checkpoint from the releases and finetune this model to new data. The second option is generally recommended, because it's faster and can get by with fewer data.
The dataset can be in any format you have available, you just need to write a function that returns a dictionary with the keys being the path to a wav file and the value for that key being the transcription of what is said in that audio. For Turkish you'd also have to add handling in the text processing, because Turkish is not yet supported. (This will change in a few versions, we will add support for a lot more languages.) Simply looking up the ID of Turkish in the espeak phonemizer (https://github.com/espeak-ng/espeak-ng/blob/master/docs/languages.md) and adding it to the TextFrontend should be enough, like in the following:
You will then also need to add a sentence for testing the learning-progress in the train loop, since it's a new language, like in the following:
To train a model with the new data, have a look at the finetuning example:
There are a bunch of comments that guide you through the changes you need to make.
AlexSteveChungAlvarez
First of all, many thanks for a great repo. I'm kind of new to this stuff, please forgive me. Can we train a new speaker and language using this repo, for example Turkish. I would be very grateful if you could provide information on what the structure of the data set should be and how it was prepared.