Musika
GAN system for infinite-length waveform music, faster than real-time on CPU.
About
Musika attacks a different corner of music generation: speed and length. The 2022 system from Marco Pasini and Jan Schlueter at JKU Linz combines hierarchical adversarial autoencoders with a latent GAN, plus a latent coordinate system that lets it generate arbitrarily long, stylistically coherent audio. Training takes about 23 hours on a single RTX 2080 Ti for hundreds of hours of music, and generation runs faster than real time on a consumer CPU.
It was demonstrated on piano music (MAESTRO) and techno, with conditioning on note density and tempo, and the paper reports lower Frechet Audio Distance than the comparable UNAGAN system. Source code and pretrained models were released, so it remains a practical option for interactive or game-audio use where latency matters more than polish.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
FAD on MAESTRO piano vs UNAGAN, lower (provider-reported).
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: ISMIR 2022 paper, open code and pretrained models, documented training cost and speed. Strengths: infinite-length generation, faster-than-real-time on CPU, single-GPU training. Docked for: (1) 2022-era GAN audio quality; (2) evaluations are provider-run. License: open-source academic release; check repo license before commercial use.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · ISMIR 2022; JKU Linz (Pasini, Schlueter), arXiv 2208.08706
- · Hierarchical adversarial autoencoders plus latent GAN
- · Infinite-length generation, faster than real-time on CPU
- · Trains on hundreds of hours with one consumer GPU
- · Code and pretrained models released; no logo staged