Music Generation

Music generation models create or modify melodies, arrangements, accompaniment, songs, and instrument sounds, often conditioned on text, audio references, symbolic scores, or visual and physiological signals. Common approaches include autoregressive Transformers and language models, diffusion and flow-matching models, neural audio codecs, and hierarchical systems that plan musical structure before rendering detailed audio. Researchers evaluate these methods on tasks such as symbolic piano generation, text-to-music, song and accompaniment generation, using benchmarks like MAESTRO, MUSDB18, and SongBench alongside listening tests. A central challenge is balancing long-range musical coherence, precise controllability, audio fidelity, and low latency, while subjective quality, cultural style, and reliable evaluation remain difficult to measure consistently.