RAVE Models by Tangible Music Lab

This is a collection of RAVE (Realtime Audio Variational autoEncoder) models trained by the Tangible Music Lab for audio generation and transformation. The aim of this repository is to provide musicians with pre-trained models for building embedded RAVE models on the Raspberry Pi platform or similar, for physical hardware and tangible interface development for sound and music experimentation. These models enable real-time audio manipulation and generation while being optimized for resource-constrained environments, making them ideal for interactive musical instruments and sound installations.

Model Description

Developed by: Tangible Music Lab
Model type: RAVE (Realtime Audio Variational autoEncoder)
License: MIT

Model Sources

Repository: https://huggingface.co/Tangible-Music-Lab/rave_models
Training Code: https://github.com/victor-shepardson/RAVE

Direct Use

These models are designed for real-time audio generation and transformation. They can be used with:

nn~
NN.ar
rave-supercollider

Models

tam_freesoundloop10k_default_b2048_r48000_z16.ts

Dataset: Freesound Loop Dataset
Model: RAVE v3 with default configuration
Latent dimensions: 16
Sample rate: 48kHz

tam_freesoundloop10k_raspi_b2048_r44100_z16.ts

Dataset: Freesound Loop Dataset
Model: Modified RAVE v3, optimized for Raspberry Pi 5
Latent dimensions: 16
Sample rate: 44.1kHz
Special features: Scaled down for real-time performance on RPi 5

Features

All models are exported for streaming inference
Compatible with nn~, NN.ar, and rave-supercollider
Models focus on encoder-decoder architecture without prior networks
Training checkpoints provided for transfer learning
For training, use the Intelligent Instruments Lab RAVE fork: https://github.com/victor-shepardson/RAVE

Training Details

Training Data

The models were trained on the Freesound Loop Dataset (FSL10K), a comprehensive collection of musical loops curated for machine learning applications. The dataset consists of 9,455 loops from Freesound.org. All sounds in the dataset are licensed under various Creative Commons licenses.

Training Procedure

Training checkpoints are provided for both models to enable transfer learning on custom datasets.

Citation

@misc {tangible_music_lab_2025,
   author       = { {Tangible Music Lab} },
   title        = { RAVE Models },
   year         = 2025,
   url          = { https://huggingface.co/Tangible-Music-Lab/rave_models },
   publisher    = { Hugging Face }
}