This repository is the vocoder of our model (HPMDubbing), which is used to convert the mel-spectrogram generated by our model into time-domain waveform.

Pretrained Models

We provide the pretrained models. One can download the checkpoints of generator (e.g., g_05000000) within the listed folders.

Folder Name	Sampling Rate	Hop Length	Segment Size	Win Length	Params.	Dataset	Fine-Tuned
HPM_Chem	16000 Hz	160	8000	640	55M	LibriTTS	No
HPM_V2C	22050 Hz	220	9900	880	58M	LibriTTS	No

Training

1. Please run

   python train_V2C_HiFiGAN.py --config config_V2C_22050Hz.json
   

   or

   python train_hifigan_16KHz.py --config config_Chem_16KHz.json
   

Inference

1. inference.py : wav -> mel -> wav

   python inference.py --checkpoint_file [Your path of checkpoint_file]
   

2. inference_e2e.py : mel -> wav

   python inference_e2e.py --checkpoint_file [Your path of checkpoint_file]
   

tensorboard

1. Please run

   tensorboard --logdir HifiGAN_16/logs/ --port=[Your port]
   

   or

   tensorboard --logdir My_vocoder_V2C/logs/ --port=[Your port]
   

References

• HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis, J. Kong, et al.

Citation