Train your own Neural Audio Model
Create custom neural audio synthesizers using RAVE, AFTER, and DDSP. No coding required.
Supported Models
Realtime Audio Variational autoEncoder by IRCAM. Perfect for fast, high-quality timbre transfer and real-time synthesis.
Audio Features Transfer and Exploration in Real-time. A diffusion-based model for precise control over timbre and structure.
Differentiable Digital Signal Processing by Magenta. Combines DSP with deep learning for realistic instrument modeling.
Why train with us?
No Code Required
Upload your audio, select a model, and let us handle the training pipeline.
Cloud GPU Training
We use powerful GPUs to train your models quickly, so you don't need expensive hardware.
Ready for Production
Export models directly for use in your DAW or live performance setup.
Frequently Asked Questions
The data requirements vary depending on the model you choose:
- DDSP: Approximately 10-15 minutes of clean, monophonic recordings. Ideally with MIDI transcription (or easy to be transcribed). Best for single instrument timbres.
- RAVE: 2-3 hours of clean, high-quality coherent audio (single style). Works with diverse sound types and can handle more complex timbres.
- AFTER: Typically more than 1 hour of audio samples for good results, ideally with MIDI transcription (or easy to be transcribed). Supports polyphonic content.
Make sure you own the rights to use the audio as training data. You agree to be the one responsible for copyright compliance in case of prejudice.
For detailed best practices, check the documentation for each model: RAVE, AFTER, DDSP.
You own the model you train, as long as you also own or have the right to use the training data. Neural Analog is the training tool; copyright responsibility for uploaded datasets stays with you.
Start from Train Neural Audio Models and check the model docs before preparing a dataset.
The models are trained using high-performance GPU pipelines. The training process includes:
- Preprocessing your audio data to extract relevant features
- Data augmentation using advanced AI models to improve robustness
- Multi-stage training with optimized hyperparameters for each model type
- Quality validation to ensure your model meets performance standards
All of this happens automatically in the cloud—you just upload your data, monitor with the available tools, and wait for the results. Start from Train Neural Audio Models.
Each model type requires its own VST/AU plugin to run in your DAW:
- RAVE: Use Neutone FX VST, IRCAM RAVE VST, or nn~ for Max/MSP and PureData
- AFTER: Max for Live devices (Ableton only) or nn~ for Max/MSP
- DDSP: Neutone FX VST or the Magenta DDSP VST
For detailed setup instructions, check the documentation for each model: RAVE, AFTER, DDSP.
Training time depends on the model complexity and size. On a single high-end GPU:
- DDSP: Approximately 6 hours
- RAVE: Approximately 12 hours
- AFTER: Approximately 24 hours
You'll receive email notifications when your model is ready for download. Training happens in the background, so you can close your browser and come back later. Start from Train Neural Audio Models when you are ready to upload a dataset.