NetEase Cloud Music's CloudEar wins MIREX 2026 music-evaluation track at 73.9%
NetEase Cloud Music's in-house CloudEar music-evaluation model took first place in the MIREX 2026 track "Music Evaluation via CMI-RewardBench" with 73.9% accuracy, according to PingWest. The track asks a model to judge which of two generated songs better matches human preference given the songs and a text prompt, scored mainly on judgement accuracy. The report says CloudEar beat the organisers' CMI-RM Baseline (65.9%) by 8 percentage points and the open SongEval baseline (72.7%) built by Northwestern Polytechnical University, Shanghai Conservatory of Music, the University of Surrey and Hong Kong University of Science and Technology. NetEase says the model, developed by its AI Music Lab, combines perception evaluation, audio-quality diagnosis and cross-modal alignment and is already used in its products; the report's framing that CloudEar is closest to human aesthetic judgement is presented as its own inference rather than an independently verified result.
The report notes that automatic preference judging is the "referee" needed to train the next generation of music models and directly sets the quality ceiling of generated music, making benchmark results in this track a reference point for AI music evaluation.
近日，2026年国际音频检索领域顶级大赛MIREX（Music Information Retrieval Evaluation eXchange）结果正式揭晓，网易云音乐智慧耳音乐评价模型CloudEar在前沿赛道“Music Evaluation via CMI-RewardBench（基于复合多模态指令的音乐评价）”以准确率73.9%的总成绩摘得冠军，并创造了该赛道世界最高纪录，这意味着网易云音乐的智慧耳音乐评价模型在判断力上与人类审美最为接近。本次赛事中，网易云音乐不仅击败众多全球顶尖高校与科技团队，还以显著优势胜过由出题方亲自搭建的官方基线模型，展现出在音乐音频领域深厚的AI技术实力。
MIREX是全球音频检索与AI音乐领域历史最悠久、最具权威性的国际评测赛事，被称为AI音乐领域“奥林匹克”，长期被全球高校与科技公司视为检验AI音乐技术水平的"黄金标尺"。它与音频检索领域的顶级国际学术会议ISMIR（International Society for Music Information Retrieval）同期举办，作为其官方评测环节。MIREX历届冠军成绩常被学界、业界直接引用为该方向的SOTA（State-of-the-Art，最先进水平）依据。
#NetEase Cloud Music#CloudEar#MIREX#ISMIR#SongEval#CMI-RM Baseline#Sony AI#Carnegie Mellon University#Georgia Institute of Technology#KAIST#Fudan University#NetEase Tianyin