ICASSP2027-AudioEditingChallenge/Track2-Agent

ICASSP 2027 Audio Editing Challenge Track 2: Agent Track

Test inputs for Track 2 of the ICASSP 2027 Audio Editing Challenge. The set has 500 examples. They cover every MMAE audio modality and every complexity category: single, multi-part, multi-instruction, multi-audio, multi-round, and multi-hop.

This release contains only input audio and instructions. The evaluation rubrics stay private until the official results are finalized.

Files

track2_agent.json     # metadata (id + messages)
wav/<id>/audioN.wav   # input audio

hf download ICASSP2027-AudioEditingChallenge/Track2-Agent --repo-type dataset --local-dir track2

Format

messages uses the same chat format as MMAE. Each audio_url is relative to the dataset root.

{
  "id": "<sample_id>",
  "messages": [
    {"role": "user", "content": [
      {"type": "text", "text": "<editing instruction>"},
      {"type": "audio", "audio_url": "wav/<sample_id>/audio1.wav"}
    ]}
  ]
}

• Multi-audio examples list several audio entries in one turn.
• Multi-round examples have several user turns. Apply them in order and submit only the final audio.

Submission

Each JSON entry is one example. Submit exactly one edited 16 kHz mono WAV per id (500 in total), plus a submission.jsonl with one line per example:

{"id": "<sample_id>", "audio_path": "audio/<sample_id>.wav"}

Package submission.jsonl together with the audio/ folder and upload it to the leaderboard. The full rules are on the Track 2 page.

Baseline

The baseline uses DeepSeek-V4-Flash as a router over DSP tools, SAM-Audio-Large, and AuK with the Prompt Enhancer. It was scored on this test set with the Qwen3-Omni-30B-A3B-Instruct judge:

IFR	CR	EMR
44.13%	76.09%	8.00%

Teams are ranked by EMR. Ties are broken by IFR, then by CR.

Contact

zhikangniu@sjtu.edu.cn, tuwenming@sjtu.edu.cn