A Python/Pytorch app for easily synthesising human voices

• Windows 10 or Ubuntu 20.04+ operating system
• 5GB+ Disk space
• NVIDIA GPU with at least 4GB of memory & driver version 456.38+ (optional)

• Automatic dataset generation (with support for subtitles and audiobooks)
• Additional language support
• Local & remote training
• Easy train start/stop
• Data importing/exporting
• Multi GPU support

• Installation
• Building the dataset
• Training
• Synthesis
• Making changes

• Add support for Talknet
• Add GTA alignment for Hifi-gan
• Improved batch size estimation
• AMD GPU support

• Try out existing voices at uberduck.ai and Vocodes
• Youtube data fetching (created by Diskr33t#5880)
• Synthesize in Colab (created by mega b#6696)
• Generate youtube transcription (created by mega b#6696)
• Wit.ai transcription

This project uses a reworked version of Tacotron2. All rights for belong to NVIDIA and follow the requirements of their BSD-3 licence.

Additionally, the project uses DSAlign, Silero, DeepSpeech & hifi-gan.

Thank you to Dr. John Bustard at Queen's University Belfast for his support throughout the project.

Supported by uberduck.ai, reach out to them for live model hosting.

Also a big thanks to the members of the VocalSynthesis subreddit for their feedback.

Finally thank you to everyone raising issues and contributing to the project.