Clone of the official FreeV, mel-to-wave vocoders with Pseudo Inversed Mel Filter.

One liner code:

model_input = mel_spec @ mel_filter.pinverse().abs().clamp_min(1e-5)

Official demo page.

git clone https://github.com/BakerBunker/FreeV.git
cd FreeV
pip install -r requirements.txt

I tried using PGHI(Phase Gradient Heap Integration) as phase spec initialization. But sadly it didn't work.

Model checkpoints and tensorboard training logs available at: huggingface

python <train-script>

Checkpoints and copy of the configuration file are saved in the checkpoint_path directory in config.json.

Modify the training and inference configuration by modifying the parameters in the config.json.

Download pretrained model on LJSpeech dataset at huggingface.

Modify the inference.py to inference.

@misc{2406.08196,
Author = {Yuanjun Lv and Hai Li and Ying Yan and Junhui Liu and Danming Xie and Lei Xie},
Title = {FreeV: Free Lunch For Vocoders Through Pseudo Inversed Mel Filter},
Year = {2024},
Eprint = {arXiv:2406.08196},
}

• official APNet2