Summary
RAVE is a realtime audio variational autoencoder developed by IRCAM designed for fast and high-quality neural audio synthesis. It is available as an open-source implementation on GitHub.
Features
- Variational autoencoder model optimized for realtime audio synthesis
- Multiple configurations available including continuous and discrete models, causal convolutions for low latency, and various architectural options
- Supports training on custom datasets with dataset preprocessing tools allowing lazy loading of raw audio files
- Augmentation techniques for improved generalization such as mute, compress, and gain adjustments
- Export models for realtime streaming with cached convolutions
- Pretrained models provided and ongoing model updates
- Integration with Max/MSP and PureData via nn~ for live audio processing and style transfer
- Batch offline audio generation support
- Open-source with GPU acceleration support (torch and torchaudio dependencies)
Usage and Community
RAVE can be used in research, music performance, and installation contexts. Users are encouraged to participate in discussion on the project Discord server and IRCAM forums. Tutorials are available for neural synthesis in DAWs, Max 8, and training custom models.
Comments
Sign in to rate, comment and bookmark.
No comments on this tool yet.