Models

MUNRA trains tiny neural networks (<50M params) to generate audio for creative music production.

Philosophy

Most audio AI optimizes for perfection. MUNRA optimizes for character:

  • Lo-fi aesthetic (tape, vinyl, imperfection)
  • Tiny models (trainable on single GPU)
  • Open source (models, weights, training process)

Planned Models

Drummer

Generates drum loops and percussion. Trained on found sounds and acoustic objects.

Work in progress — ~20M params

Voice Reciter

Generates spoken word and vocal textures. Trained on recitations and phonemes.

Work in progress — ~28M params

Drone Generator

Generates pads and evolving textures. Trained on synths and guitar feedback.

Work in progress — ~22M params

Follow Progress

Each model's development is documented in experiments.

View experiments →