AI
New Expert-Labeled Dataset and AI Framework Boost Music Emotion Recognition Accuracy
Researchers release Memo2496, a dataset of nearly 2,500 songs annotated by music specialists, paired with DAMER, a dual-view framework that outperforms existing approaches at identifying emotional content in instrumental music.
Music emotion recognition—the task of automatically detecting how a piece of music makes listeners feel—has long struggled with sparse expert annotations and inconsistent results across different datasets. A new paper on arXiv addresses these limitations by introducing both a carefully curated dataset and a novel machine-learning system designed to improve accuracy across multiple music collections.
The Dataset Problem#
Existing music emotion datasets suffer from limited expert labeling and difficulty generalizing to unfamiliar music catalogs. Researchers created Memo2496, a collection of 2,496 instrumental tracks where emotions are mapped along two dimensions: valence (how happy or sad) and arousal (how calm or energetic). Each track received continuous ratings from 30 certified music specialists who underwent training and recalibration to ensure consistency.
How DAMER Works#
The team developed the Dual-view Adaptive Music Emotion Recogniser (DAMER), a framework combining three key innovations:
-
Dual-Stream Attention Fusion (DSAF): Processes two complementary audio representations—Mel spectrograms and cochleagrams—simultaneously, allowing the model to capture both traditional frequency patterns and perceptual aspects of how the human ear processes sound.
-
Progressive Confidence Labelling (PCL): Uses a curriculum-learning approach where the system gradually refines its predictions by adjusting temperature parameters and measuring divergence, mimicking how humans learn incrementally.
-
Style-Anchored Memory Learning (SAML): Maintains a memory bank of emotional embeddings that helps the model recognize similar emotions even when songs sound acoustically different.
Performance and Validation#
When tested on Memo2496 alongside two widely-used external datasets (1000songs and PMEmo), DAMER achieved the strongest results for arousal detection on both Memo2496 and 1000songs, and the highest valence accuracy on PMEmo. The researchers validated each component through ablation studies and released both the dataset and source code publicly, enabling reproducibility and future research.
