Reproduce “LiveBand: Live Accompaniment Generation in the Audio Domain” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Marco Pasini, Javier Nistal, Ben Hayes, Mathias Rose Bjare, Stefan Lattner, George Fazekas
LiveBand is a real-time system that generates high-fidelity music accompaniments for live audio input, overcoming the need for future input lookahead and eliminating exposure bias through its design. This system consistently outperformed previous autoregressive models like StreamMusicGen across objective metrics and was subjectively preferred for audio quality, consistency, and mix adherence, achieving a quality score of 2.6 vs. SMG's 1.9 and a mix adherence score of 3.4 vs. SMG's 2.2 in listening tests.
Key insights
- Robust, mix-adherent real-time accompaniment generation is possible without relying on future input lookahead, disproving prior assumptions.
- Intrinsic alignment of training and inference conditions in the generator design effectively eliminates exposure bias, leading to stable, long-term generative performance.
- Sequence-level adversarial supervision proves more effective for real-time music generation than frame-level objectives, fostering holistic musical coherence and allowing for natural timing nuances.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.