NetEase Cloud Music's CloudEar Model Wins MIREX 2026 Music Evaluation Track

The results of MIREX 2026, a leading international evaluation in audio retrieval and AI music research, were announced in October 2026. NetEase Cloud Music's self-developed music evaluation model CloudEar won first place in the music evaluation track with an accuracy of 73.9%, setting a new track record. The model outperformed teams from Georgia Tech, KAIST, Carnegie Mellon University, Sony AI and Fudan University, as well as the official baseline model.

MIREX is held alongside ISMIR and is widely regarded as a benchmark for AI music research. In this year's music evaluation task, models were asked to judge which of two generated songs better matches human preferences using the generated tracks and text prompts. CloudEar uses a collaborative architecture of three expert components: perceptual evaluation, sound quality diagnosis and cross-modal alignment, combining pre-trained music representations, multi-task preference learning, DSP-based audio features and text-music semantic matching. Its 73.9% accuracy surpassed the official CMI-RM Baseline (65.9%) by 8 percentage points and also exceeded the open academic baseline SongEval (72.7%).

CloudEar was developed by NetEase Cloud Music's AI Music Lab and has been applied in the company's products for AI music quality control, content distribution and generative model optimization. The lab previously won MIREX titles in 2020 and 2021.