The 24kHz model was pretrained using LibriTTS dataset with a full 100-band mel spectrogram as input (see
config.json for the exact hyperparameter setup) with the BigVGAN
repository. The pretraining was performed over 1300k steps with a 100 batch size with 8 A100 40GB GPUs.
The run the inference with the example command for generating audio from the model. It computes mel spectrograms using wav files from --input_wavs_dir and saves the generated audio to --output_dir.
python NEMO_PATH/inference.py \
--checkpoint_file MODEL_PATH/BigVGAN-L/g_01300000.pt \
--input_wavs_dir AUDIO_PATH/input_wav \
The vocoder can be finetuned further on using the NEMO_PATH/train.py script as the checkpoints save all the optimizer information.
- Downloads last month