train_coqui_tts_ita

My guide to create an italian TTS with Coqui

Prepare the enviroment

• clone this repo and go with a terminal inside
• create a venv

python3.8 -m venv venv
source ven/bin/activate
# or for fish shell
source ven/bin/activate.fish

• clone the coqui repo (I use c63bb481e95bd4a1ff978947d8e3e6c0bfb4177f sha version) and install all libs

git clone https://github.com/coqui-ai/TTS.git
cd TTS
pip install -r requirements.txt
python setup.py install
cd ..

Train the model

• modify the train_glowtts.py with the correct dataset_path and output_path, eventually change the batch size if you do not have nought ram into GPU

# activate venv
PYTORCH_CUDA_ALLOC_CONF="max_split_size_mb:25" CUDA_VISIBLE_DEVICES=1 python src/train_glowtts.py

Use trained model

• predict one text from commandline

CUDA_VISIBLE_DEVICES=0 tts --text "Ciao Pippo" --model_path "OUTPUT_PATH/best_model.pth.tar" --config_path "OUTPUT_PATH/config.json"

• run a prediction server to play with the trained model

CUDA_VISIBLE_DEVICES=0 tts-server --model_path "OUTPUT_PATH/best_model.pth.tar" --config_path "OUTPUT_PATH/config.json"

Results

Model	Dataset
glowtts	male-LJSpeech-italian
glowtts	female-LJSpeech-italian
vits	male-LJSpeech-italian
vits	female-LJSpeech-italian