Pre-requisites

1. Python == 3.8
2. Clone this repository.
3. Install python requirements. Please refer requirements.txt
4. Download a 48k dataset, such as genshin or VCTK.

SR model sample theory

SR results

• origin 16k mel-spectrum
• generated 48k mel-spectrum

Training

CUDA_VISIBLE_DEVICES="3" python train.py \
--config config_v2_16k_to_48k.json \
--input_wavs_dir VCTK-Corpus/wav48/,genshin --checkpoint_path exp/v2_16k_to_48k/

To train 24k_to_48k, replace config_v2_16k_to_48k.json with config_v2_24k_to_48k.json.
Checkpoints and copy of the configuration file are saved in checkpoint_path directory by default.
You can change the path by adding --checkpoint_path option.

1. The hifigan training speed may be a bit slow. You can use HiFTNet-sr to accelerate training, and its generate result may be sounds better.

Pretrained Model

1. The pretrained models provided is in "exp/v2_16k_to_48k/g_00120000", trained with StarRail_Datasets and VCTK.
2. For i don't have GPU resources, a kind person(@Lucy) train config_v2_16k_to_48k version and trained stop at 120k, not trained enough. So the results maybe sounds a little electronic, but verified that model methods is valid.
3. You can use other hifigan config version to train, or use HiFTNet-sr.

Inference from wav file

1. Make test_files directory and copy wav files into the directory.
2. Run the following command.

   # python inference.py --checkpoint_file [generator checkpoint file path]
   CUDA_VISIBLE_DEVICES="-1" python inference.py --checkpoint_file exp/v1_16k_to_48k/g_00120000 --input_wavs_dir LJSpeech-BZNSYP-16k
   

Generated wav files are saved in generated_files by default.
You can change the path by adding --output_dir option.

Reference

Our repository is heavily based on jik876 's hifi-gan.