ICASSP2027-AudioEditingChallenge/Track1-Single-Model

ICASSP 2027 Audio Editing Challenge Track 1: Single Model Track

Test inputs for Track 1 of the ICASSP 2027 Audio Editing Challenge. The set has 500 examples. All of them are in the MMAE single complexity category and cover speech, music, sound, and their mixtures.

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

Files

track1_single_model.json   # metadata (id + messages)
wav/<id>/audio1.wav        # input audio

hf download ICASSP2027-AudioEditingChallenge/Track1-Single-Model --repo-type dataset --local-dir track1

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"}
    ]}
  ]
}

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 1 page.

Baseline

The baseline is the AuK base model with the Prompt Enhancer disabled. It was scored on this test set with the Qwen3-Omni-30B-A3B-Instruct judge:

IFR	CR	EMR
40.33%	74.70%	7.60%

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

Contact

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