arXiv:2611.14724 · v1

Harmony: Bridging Audio Understanding and Editing through Agentic Collaboration

首次提交 2026-11-11 04:36

版本更新 2026-11-11 04:36

雷达观测 2026-11-13 10:09

摘要

Natural-language audio editing requires connecting a user's intent to specific sounds and time regions while preserving surrounding content. Complex requests make this connection demanding: an edit may select one occurrence, depend on a temporal relation, or modify a source that overlaps with other sounds. We introduce Harmony, an agentic framework that bridges audio understanding and editing through collaboration among an LLM planner, an audio-language model, and external tools. Harmony combines two mechanisms. Edit-driven collaborative planning uses targeted questions and tool measurements to determine the target, timing, and operation parameters. Evidence-guided execution uses observations of the first separated or generated audio to trigger bounded corrections before final mixing. Together, these mechanisms connect semantic decisions to waveform operations throughout the editing process. Harmony supports addition, removal, replacement, and volume adjustment. On 420 MMAE examples spanning these operations, Harmony achieves an exact match rate of 22.62%, compared with 13.57% for the strongest baseline, and improves aggregate instruction following and context preservation.