MusiChat presents a conversational music-authoring system for iteratively refining an evolving composition through natural language, addressing the limitations of prompt-and-regenerate workflows. Its hierarchical controllable generation framework separates lyric-aligned structural generation from expressive surface realization, enabling style changes and structure-preserving edits. A memory-augmented architecture tracks the active composition and interaction history, while hybrid intent routing handles both precise musical edits and open-ended creative requests. The paper reports 95.31% single-turn and 100% multi-turn interaction accuracy, plus like-to-dislike ratios of 2:1 for melody naturalness and 3:1 for overall musical quality.
No heat snapshots are available in the last 24 hours.