A codebase for Kinyarwanda speech synthesis (text-to-speech) based on MB-iSTFT-VITS2 model in PyTorch. The original codebase and description comes from MB-iSTFT-VITS2 which itself is a hybrid combination of vits2_pytorch and MB-iSTFT-VITS.
An architectural depiction of the model is presented below and its details can be found the original sources.
- Checkout the code in the Inference directory and install
monotonic_alignthekinyattsmodules
pip install -e ./Inference/monotonic_align/
pip install -e ./Inference/- Download the pre-trained Kinyarwanda TTS model TTS_MODEL_ms_ktjw_istft_vits2_base_1M.pt
- Go to the
kinyattssub-directory and run an inference server using uwsgi
cd Inference/kinyatts/
nohup sh run.sh &
- Alternatively use the provided Jupiter notebook to synthetise speech: Inference/kinyatts/kinyatts_inference.ipynb
Follow instructions in Training codebase to install requirements and train a basic multi-speaker TTS model.