EmoMusic: Dimensional MER Benchmark

Updated 14 July 2026

EmoMusic is a music emotion recognition dataset defined by continuous valence–arousal ratings on 744 diverse, 45-second audio clips.
It supports regression-based MER research through static label splits and integration into multitask learning frameworks with feature fusion and knowledge distillation.
Studies using EmoMusic highlight challenges in cross-dataset generalization due to genre diversity and heterogeneous annotation schemes.

Searching arXiv for papers mentioning EmoMusic and related MER benchmarks. Searching arXiv for "EmoMusic music emotion recognition". EmoMusic is a music emotion recognition dataset and benchmark centered on continuous valence–arousal annotation. It was developed for the “Emotion in Music” MediaEval 2013 task, contains 744 audio clips of 45 seconds each, spans styles such as Country, Blues, Electronic, and Rock, and uses crowd-sourced annotations of perceived emotion on continuous valence and arousal scales ranging from $-1$ to $+1$ . In subsequent work, it is treated as a dimensional-label benchmark with static valence and arousal values per music excerpt and as a recurring test bed for representation learning, multitask training, and cross-dataset generalization in Music Emotion Recognition (MER) (Ching et al., 6 Oct 2025, Kang et al., 6 Feb 2025).

1. Dataset identity and research role

EmoMusic is situated within the strand of MIR research that models musical affect in a continuous space rather than through discrete mood tags. The dataset was introduced for the “Emotion in Music” MediaEval 2013 task, and its annotations focus exclusively on perceived emotion rather than induced emotion. In the comparative characterization reported later, EmoMusic is also described as typically comprising commercial music, with listener annotations and audio-plus-lyrics data, which distinguishes it from later songwriter-centric or MIDI-enriched resources (Ching et al., 6 Oct 2025, Hwang et al., 6 Nov 2025).

Its continuing relevance comes from the fact that it provides a compact but nontrivial dimensional benchmark. One later framework explicitly states that EmoMusic uses static valence and arousal values per music excerpt, with $N=744$ , and had to be randomly split into 70% train, 15% validation, and 15% test because of the lack of official splits (Kang et al., 6 Feb 2025). This combination of fixed continuous labels, modest scale, and stylistic variety has made it particularly useful for testing whether MER models actually learn transferable affective structure rather than merely memorizing dataset-specific regularities.

2. Dimensional labeling and its methodological consequences

A central methodological issue in MER is the coexistence of heterogeneous annotation schemes. Some datasets use categorical labels such as “happy” or “sad,” while others, such as EmoMusic, use dimensional labels in valence and arousal. The unified MER framework of 2025 treats EmoMusic alongside DEAM and PMEmo as a regression task, while datasets such as MTG-Jamendo are handled through a separate multi-label classification branch (Kang et al., 6 Feb 2025).

This dimensional formulation contrasts sharply with categorical alternatives such as GEMS-9, where nine emotion categories—amazement, solemnity, tenderness, nostalgia, calmness, power, joyful activation, tension, and sadness—are modeled as a multilabel and multiclass problem on the Emotify dataset (Paolizzo et al., 2019). The contrast is consequential: categorical MER emphasizes label co-occurrence structure, whereas EmoMusic is used to estimate continuous affect coordinates. A plausible implication is that EmoMusic has become especially important whenever the research question concerns regression, cross-dataset calibration, or compatibility with Russell-style circumplex formulations.

The dimensional regime also affects evaluation culture. In the EmoMusic-focused multitask framework, performance is reported with the coefficient of determination for valence and arousal, $R^2_V$ and $R^2_A$ , rather than category-wise accuracy or multilabel metrics (Kang et al., 6 Feb 2025). This places EmoMusic in the lineage of VA-based affective MIR rather than music tagging.

3. Use in unified MER architectures

A prominent recent use of EmoMusic is the unified multitask learning framework that combines categorical and dimensional labels across datasets. The architecture has two output branches: a categorical branch for multi-label classification and a dimensional branch for continuous valence–arousal regression. Only the relevant branch is updated for each batch; for EmoMusic, only the dimensional branch and its loss are updated. The model combines MERT embeddings with musical features derived from chord progressions and key signatures, and it augments supervised learning with knowledge distillation from teacher models trained on individual datasets (Kang et al., 6 Feb 2025).

For EmoMusic, the regression loss is combined with distillation as

$L_{\text{Total,VA}} = \beta \cdot L_{\text{VA}} + (1 - \beta) \cdot L_{\text{KD}},$

with $\beta = 0.2$ in the reported setup (Kang et al., 6 Feb 2025).

The reported EmoMusic results show how feature fusion and cross-dataset training alter performance.

Configuration	$R^2_V$	$R^2_A$
Only MERT features	0.5657	0.7404
MERT + Chord/Key	0.5957	0.7489
Full multitask/KD regime	0.6512	0.7616

These numbers indicate two distinct effects. First, adding chord and key features improves on MERT-only embeddings. Second, training jointly on multiple datasets with heterogeneous label types improves further. In the formulation of that work, EmoMusic therefore functions not only as a benchmark but also as evidence that dimensional MER can benefit from shared latent structure learned from categorical corpora (Kang et al., 6 Feb 2025).

4. Distribution gap and out-of-distribution generalization

A major contemporary reinterpretation of EmoMusic comes from work on the data distribution gap in MER. That study uses five datasets with dimensional emotion annotations—EmoMusic, DEAM, PMEmo, WTC, and WCMED—and shows that strong in-dataset performance on EmoMusic does not imply strong cross-dataset generalization. EmoMusic is described there as more genre-diverse than WTC or WCMED, with official genre labels indicating dominance by popular genres, although the authors note that “Classical” labels in EmoMusic may be unreliable (Ching et al., 6 Oct 2025).

Using Jukebox embeddings, the reported average $R^2$ values make the out-of-distribution problem explicit.

Training $+1$ 0 Test	Avg. $+1$ 1
EmoMusic $+1$ 2 EmoMusic	0.674
EmoMusic $+1$ 3 DEAM	0.44
EmoMusic $+1$ 4 PMEmo	-0.32
EmoMusic $+1$ 5 WTC	0.05
EmoMusic $+1$ 6 WCMED	-0.26

The same study reports that deep audio embeddings tend to cluster strongly by dataset and genre, and that annotation normalizations to $+1$ 7 can mask genuine differences in emotional content between datasets. For EmoMusic versus WCMED, the data feature distribution gap is large— $+1$ 8, $+1$ 9—while the annotation gap is small— $N=744$ 0, $N=744$ 1 (Ching et al., 6 Oct 2025). This directly challenges a common misconception that a shared valence–arousal scale is sufficient to make MER datasets interchangeable.

The proposed remedy in that paper is deliberately simple: combine Jukebox embeddings with chroma features and train on several stylistically diverse datasets. On WCMED, an EmoMusic-only MLP with Jukebox features yields $N=744$ 2; replacing the input with Jukebox+Chroma raises it to $N=744$ 3; and multi-dataset training on EmoMusic+PMEmo+WTC with Jukebox+Chroma reaches $N=744$ 4. On the combined test set, the same multi-dataset Jukebox+Chroma setting reaches $N=744$ 5 (Ching et al., 6 Oct 2025). This suggests that EmoMusic is valuable not simply as a standalone benchmark, but as one component of a deliberately heterogeneous training mixture.

5. Relation to adjacent emotion datasets

Later datasets clarify what EmoMusic does and does not provide. Emo100DB is a dataset of 100 improvised songs by 20 young adults, each with lyrics, transcribed melody in MIDI, original WAV audio, trimmed WAV segments, and songwriter-provided quadrant annotations based on Russell’s circumplex model. In the comparison given there, EmoMusic is characterized as commercial music with listener annotations, audio and lyrics, and listener-centric continuous emotion labels (Hwang et al., 6 Nov 2025).

Memo2496 extends the continuous VA tradition in a different direction. It offers 2,496 instrumental tracks annotated by 30 certified music specialists, with annotation quality controlled through calibration with extreme emotion exemplars and an intra-annotator Euclidean V–A distance threshold of 0.25. Its associated DAMER framework then uses Mel spectrograms and cochleagrams, Progressive Confidence Labelling, and Style Anchored Memory Learning to address pseudo-label reliability and cross-track feature drift (Li et al., 16 Dec 2025). Relative to EmoMusic, this represents a shift toward larger scale, expert annotation, and more explicit control of annotation consistency.

These later resources do not replace EmoMusic so much as reframe it. EmoMusic remains one of the canonical dimensional benchmarks, but subsequent datasets expose axes along which it can be complemented: symbolic access, songwriter-intent labels, instrumental-only design, and stricter annotation governance.

6. Broader significance in affective music computing

The importance of EmoMusic extends beyond benchmark reporting because its dimensional valence–arousal formulation aligns with a broad range of affective music systems. Interactive emotion manipulation of songs has been formulated through Russell’s circumplex model, with four emotional quadrants and explicit projection of classifier outputs to 2D affective coordinates (Abdalla et al., 2024). Portable EEG–fNIRS studies of music-induced affect likewise operationalize valence and arousal as binary targets and validate emotion induction with AI-generated music stimuli (Zhao et al., 5 Aug 2025). Video-based music generation has gone further by conditioning a MIDI transformer on continuous emotional values rather than discrete categories, mapping multimodal video emotion predictions into final valence and arousal through mixture means (Sulun, 5 Feb 2026).

Within that broader landscape, EmoMusic occupies a specific epistemic niche: it is a listener-annotated, perceived-emotion, continuous-label dataset that has repeatedly been used to test whether MER models can regress affective coordinates from audio. Its main scholarly significance lies in how clearly it exposes the central tensions of MER: dimensional versus categorical emotion models, in-domain accuracy versus cross-dataset generalization, and shared affective representation versus genre- or dataset-specific bias.

A balanced reading is therefore that EmoMusic is neither a complete proxy for musical affect nor an obsolete benchmark. Rather, it is a compact, historically important dimensional corpus whose value now lies as much in what it reveals about transfer failure, label heterogeneity, and benchmarking practice as in the absolute performance numbers it enables (Kang et al., 6 Feb 2025, Ching et al., 6 Oct 2025).