Light Speed ⚡ is an open-source text-to-speech model based on VITS, with some modifications:
- utilizes phoneme duration's ground truth, obtained from an external forced aligner (such as Montreal Forced Aligner), to upsample phoneme information to frame-level information. The result is a more robust model, with a slight trade-off in speech quality.
- employs dilated convolution to expand the Wavenet Flow module's receptive field, enhancing its ability to capture long-range interactions.
We provide two pretrained models and demos:
- VN - Male voice: https://huggingface.co/spaces/ntt123/Vietnam-male-voice-TTS
- VN - Female voice: https://huggingface.co/spaces/ntt123/Vietnam-female-voice-TTS
Q: How do I create training data?
A: See the ./prepare_ljs_tfdata.ipynb
notebook for instructions on preparing the training data.
Q: How can I train the model with 1 GPU?
A: Run: python train.py
Q: How can I train the model with 4 GPUs?
A: Run: torchrun --standalone --nnodes=1 --nproc-per-node=4 train.py
Q: How can I train a model to predict phoneme durations?
A: See the ./train_duration_model.ipynb
notebook.
Q: How can I generate speech with a trained model?
A: See the ./inference.ipynb
notebook.
- Most of the code in this repository is based on the VITS official repository.