EDM-98: EDM Structure Segmentation Dataset

• EDM-98 is a dataset for music structure segmentation in EDM that uses an EDM-specific taxonomy defined by energy, rhythm, and timbral changes.
• The dataset comprises 98 tracks, balanced across BPM ranges from EDM-CUE, and designed to correct the pop-centric bias of traditional datasets.
• EDM-98 underpins the EDMFormer model, which outperforms pop-centric approaches by significantly improving boundary detection and per-frame accuracy.

EDM-98 is a dataset for music structure segmentation in Electronic Dance Music (EDM), released with the transformer-based EDMFormer model as “a group of 98 professionally annotated EDM tracks.” It was introduced to address a specific failure mode of existing music structure analysis pipelines: most prior datasets and models are pop-centric and rely on lyrical or harmonic similarity, whereas EDM structure is defined more strongly by changes in energy, rhythm, and timbre, with sections such as buildup, drop, and breakdown (Sajeer et al., 8 Mar 2026).

1. Motivation and research context

The dataset was constructed in response to the observation that existing music structure analysis datasets and models are pop-centric and therefore fail to capture the energy- and rhythm-driven organization of EDM. In the formulation used alongside EDMFormer, the central problem is not merely boundary detection in generic audio, but structural modelling under a genre-specific regime in which section identity is tied to energy profiles, rhythmic transitions, and timbral change rather than verse–chorus recurrence (Sajeer et al., 8 Mar 2026).

This motivation determines both the dataset design and the annotation ontology. Traditional pop-oriented labels such as verse, chorus, and bridge were treated as inadequate for EDM. Instead, the dataset introduces an EDM-specific taxonomy intended to reflect how EDM organizes tension, release, contrast, and closure. A plausible implication is that EDM-98 is best understood not as a simple benchmark subset, but as a corrective instrument for genre mismatch in structural audio modelling.

2. Corpus construction and sampling design

EDM-98 was selected from the EDM-CUE corpus, which contains 4,710 songs. The construction pipeline first excluded all tracks without a reliably extracted BPM. The remaining tracks were then grouped into five BPM buckets covering 120–180 BPM, specifically to counter the original tempo imbalance of EDM-CUE, which was heavily skewed toward high-tempo categories such as Drum & Bass and Jungle. Sampling was randomized and performed without replacement, with an equal or as-even-as-possible number of tracks drawn from each tempo bucket and with residual assignment to achieve a final total of $N=98$ (Sajeer et al., 8 Mar 2026).

The resulting dataset is explicitly described as a balanced, stratified sample rather than a proportional subset of EDM-CUE. This distinction is methodologically important because EDM-CUE is reported as 90.5% high-tempo Drum & Bass/Jungle with mean BPM 169.52, whereas EDM-98 was designed to provide broader structural and tempo coverage suitable for evaluation and training. The final tempo statistics for EDM-98 are mean BPM 137.31, median BPM 135.00, standard deviation 20.04, and range 100–175 BPM (Sajeer et al., 8 Mar 2026).

Genre family	Share
House/Deep House	19.4%
Electro/Dance	14.3%
Techno	14.3%
Trance/Progressive	14.3%
Dubstep/Bass	14.3%
Drum & Bass/Jungle	14.3%
Chill/Downtempo	9.2%

This marginal tempo distribution was intentionally made close to uniform across mid/high-tempo genres. In context, that design choice makes EDM-98 a sampling intervention as well as an annotation resource.

3. Annotation protocol and structural taxonomy

EDM-98 contains 98 tracks annotated by two professional annotators. Segment boundaries were marked with $\pm 0.5$ seconds precision, and each segment was assigned a section label from the EDM-specific taxonomy. For this proof-of-concept version, however, each track was labeled by a single annotator, a choice explicitly acknowledged as a potential bias and one that motivated plans for inter-annotator reliability in future work (Sajeer et al., 8 Mar 2026).

The taxonomy comprises seven labels, each defined by structural function rather than pop-song form:

• Intro: Beginning, low energy, sparse instrumentation.
• Build-up: Rising energy/tension, often with increasing drum/riser patterns.
• Drop: Peak energy, main rhythm and bassline.
• Breakdown: Lower energy, often melodic/atmospheric, provides contrast.
• Outro: Ending portion, gradual energy decrease.
• Silence: Minimal or no audio signal.
• End: Explicit track ending or final note.

The paper further notes that annotation was guided by energy and rhythmic features rather than harmonic or lyrical structure. This suggests that the dataset encodes a theory of form tailored to EDM production practice: sections are defined by functional energy states and transitions, not only by recurrent melodic or textual content.

4. Formal role in EDMFormer

EDM-98 was introduced together with EDMFormer, a genre-adapted transformer for structure segmentation. The model uses self-supervised embeddings from two audio foundation models, MuQ and MusicFM, which are fused and fed into the SongFormer encoder. In the reported system, all training and fine-tuning in the final stage were conducted on EDM-98, while SongFormer was evaluated in a zero-shot setting on the same dataset (Sajeer et al., 8 Mar 2026).

The paper states the contrast between a general, pop-centric mapping and an EDM-specialized mapping as

$$f_{general}: E_{ssl}(X) \xrightarrow{\text{pop-centric dataset + general taxonomy}} (T, Y),$$

and

$$f_{edm}: E_{ssl}(X) \xrightarrow{\text{EDM-98 dataset + EDM taxonomy}} (T, Y),$$

where $E_{ssl}(X)$ denotes self-supervised audio embeddings, $T$ timestamps, and $Y$ section labels.

The evaluation protocol used an 88/10 train/test split on EDM-98 and 5-fold cross-validation for hyperparameter tuning. Metrics were boundary Hit Rate at two tolerances, [email protected] and HR@3s, together with per-frame accuracy (ACC). Within this design, EDM-98 functions simultaneously as a fine-tuning corpus, an evaluation benchmark, and an empirical test of whether genre-specific structure priors improve segmentation.

5. Reported benchmark results

The benchmark reported for EDM-98 compares SongFormer under a pop taxonomy with EDMFormer under the EDM taxonomy. The results are as follows (Sajeer et al., 8 Mar 2026):

Metric	SongFormer (Pop Taxonomy)	EDMFormer (EDM Taxonomy)
[email protected]	0.569	0.616
HR@3	0.608	0.635
ACC	0.148	0.883

The paper states that EDMFormer outperformed SongFormer across all metrics, with a 4.7% improvement at [email protected], a 2.7% improvement at HR@3s, and a 73.5% improvement in ACC. It also reports that the improvement was particularly notable for drops and buildups (Sajeer et al., 8 Mar 2026).

The strongest interpretive claim attached to these numbers is that the large increase in per-frame accuracy demonstrates a mismatch between pop-centric taxonomy and EDM structure. More cautiously, the results indicate that section labeling is highly sensitive to label ontology, not only to model capacity. This makes EDM-98 significant as an annotation framework as much as as a dataset in the narrow sense.

6. Limitations, availability, and significance

The paper presents EDM-98 as a proof-of-concept resource and states several limitations directly. First, 98 tracks is small for neural modelling. Second, each track was labeled by a single annotator, so inter-annotator agreement is not yet available. Third, dataset growth is constrained by the labor-intensive nature of high-precision structural labeling. The paper also notes that explicit segment count histograms, per-label frequency tables, and a direct dataset URL are not provided in the manuscript (Sajeer et al., 8 Mar 2026).

At the same time, EDM-98 is positioned as the first professionally annotated, tempo- and genre-balanced dataset designed explicitly for structure analysis of Electronic Dance Music. Its significance lies in demonstrating that combining learned representations with genre-specific data and structural priors is effective for EDM and could be applied to other specialized music genres or broader audio domains. On that reading, EDM-98 is not merely a benchmark of 98 tracks, but a case study in how dataset design, annotation ontology, and model specialization interact in modern music structure segmentation research (Sajeer et al., 8 Mar 2026).