--- title: README emoji: 🐨 colorFrom: gray colorTo: yellow sdk: static pinned: false ---
Advancing precise, general-purpose, instruction-based audio editing
## About the Challenge The Audio Editing Challenge at ICASSP 2027 brings together researchers working on general-purpose audio editing. Built on **MMAE (A Massive Multitask Audio Editing Benchmark)**, the challenge asks systems to understand source audio, follow potentially complex instructions, and produce precise edits while preserving unrelated content and audio quality. Tasks span **speech, music, sound, and their mixtures** — from removing applause without changing the speaker to replacing a lyric while preserving the singer and accompaniment. ## Challenge Tracks | Track | Goal | | --- | --- | | **Single Model Track** | Build one end-to-end model that directly transforms input audio according to an instruction. | | **Agent Track** | Build an autonomous system that can plan, use locally deployed models or signal-processing tools, inspect intermediate results, and refine its output. | **Track details:** [Single Model Track](https://audio-editing-challenge.github.io/track1/) and [Agent Track](https://audio-editing-challenge.github.io/track2/). Both tracks cover speech, music, sound, and their mixtures. The Single Model Track is restricted to the MMAE **single** complexity category; the Agent Track covers **all complexity categories**. ## Baselines Our **[baseline repository](https://github.com/Audio-Editing-Challenge/Audio-Editing-Challenge-Baseline)** provides setup instructions, inference scripts, and submission-format outputs for both tracks: - **Single Model:** AuK-based end-to-end editing with Prompt Enhancer disabled. - **Agent:** An LLM router combining DSP tools, SAM-Audio source separation, and AuK generative editing. For participation rules, test-set release details, and submission instructions, please follow the **[official website](https://audio-editing-challenge.github.io/)**. ## Evaluation The two tracks are ranked independently using rubric-based evaluation: - **Instruction Following Rate (IFR):** Were the requested edits correctly executed? - **Consistency Rate (CR):** Were unrelated audio content and quality preserved? - **Exact Match Rate (EMR):** Were all instruction-following and consistency rubrics satisfied for a sample? Rankings are determined by **Overall EMR**, with ties broken by Overall IFR, then Overall CR. See the [evaluation protocol](https://audio-editing-challenge.github.io/#benchmark-and-evaluation-protocol) for full details. ## Key Dates | Milestone | Date | | --- | --- | | Challenge test sets and submission SDK released; leaderboard opens | November 10, 2026 | | Final submissions due; leaderboard freezes | November 25, 2026 | | Final rankings and invited teams announced | December 8, 2026 | | Invited two-page ICASSP papers due | January 7, 2027 | Submission deadlines are **11:59 PM U.S. Pacific Time**. The schedule is tentative; consult the [official timeline](https://audio-editing-challenge.github.io/#timeline) for updates. ## Contact & Community For challenge questions and invitations to the Slack workspace or WeChat group, contact [Zhikang Niu](mailto:zhikangniu@sjtu.edu.cn) or [Wenming Tu](mailto:tuwenming@sjtu.edu.cn). The challenge prize pool is sponsored by **Tencent**. The top three teams in each track will be invited to submit a two-page ICASSP 2027 paper and present their work in the Grand Challenge session.