F0-AUTOVC: F0-Consistent Many-to-Many Non-Parallel Voice Conversion via Conditional Autoencoder

This repository provides a PyTorch implementation of the paper F0-AUTOVC.

Based on

• https://github.com/auspicious3000/autovc
• https://github.com/auspicious3000/SpeechSplit
• https://github.com/christopher-beckham/amr

Dependencies

• Python 3.7
• Pytorch 1.6.0
• TensorFlow
• Numpy
• librosa
• tqdm

Usage

1. Prepare dataset
   we used the VCTK dataset as used in original paper.
   But, you can use your own dataset.

2. Prepare the speaker to gender file as shown in nikl_spk.txt and run make_spk2gen.py

   • Format
     speaker1 gender1
     speaker2 gender2

   • Example:
     p225 W
     p226 M
     p301 W
     p302 W
     .
     .

3. Preprocess data using preprocess.py

4. Run task_launcher.py