Text-to-Music Generation

Text-to-music generation studies how to turn language descriptions into music, from audio tracks and songs to symbolic scores. Systems commonly use autoregressive models over audio tokens, diffusion or flow models over continuous latents, or hybrids that plan melodies and song sections before rendering audio. Text-audio alignment, MIDI supervision, and local prompts help control attributes such as instrumentation, lyrics, timing, and long-range structure. Evaluation combines similarity metrics, attribute-specific tests, and listening preferences; a central challenge is balancing prompt fidelity and controllability with coherent, high-quality music, while making evaluation and training-data provenance more reliable.