Raveform: A Dataset of Metrical and Functional Structure Annotations for EDM Tracks in DJ Mixes

2.1 DJ mix datasets

Only a few datasets containing DJ mixes are currently available. Sonnleitner et al. (2016) introduced two datasets for evaluating audio fingerprinting in DJ mixes and publicly released one of them. This dataset includes track lists and the corresponding tracks played in the mixes, similar to some aspects of our proposed dataset. However, its scale is limited to only 10 DJ mixes and 118 tracks, making it insufficient for large‑scale analysis. Another dataset, UnmixDB, was introduced by Schwarz and Fourer (2019). Although it also consists of DJ mixes and associated tracks, the mixes are automatically generated, which limits their suitability for musicological studies that aim to analyze human DJ behavior.

2.2 Mix/cue point datasets

The term ‘cue point’ is inherently ambiguous. While it generally refers to a notable position within a track, a ‘mix point’ specifically denotes the start or end of a transition between two tracks, resulting in four distinct positions per transition. In this article, we distinguish clearly between these two terms. Prior research on mix/cue point detection is related to three main tasks: (1) mix segmentation that divides a DJ mix into its constituent tracks, (2) mix/cue point extraction that identifies such points given a DJ mix and track list, and (3) mix/cue point estimation that automatically predicts transition points for a given track. Studies on mix segmentation (Glazyrin, 2014; Scarfe et al., 2014) can detect transitions between tracks but are limited in semantic interpretation due to the lack of full track context. In cue point extraction, Schwarz and Fourer (2019) proposed a method evaluated only on synthetic data. Kim et al. (2020) developed a method for mix point extraction using a real‑world dataset (1.5k DJ mixes, 15k tracks), but the dataset is not publicly available and lacks structural annotations, limiting deeper analysis. For cue point estimation, several heuristic approaches exist (Schwarz et al., 2018; Veire and De Bie, 2018), but they are not data‑driven. Recently, Argüello et al. (2024) have released a dataset containing 4,710 tracks and 21k cue points (on average, 4.6 per track) curated from four professional DJs, along with a data‑driven estimation model. However, the inherent ambiguity of the term ‘cue point’ remains problematic. Annotators often apply inconsistent criteria, and a cue point generally indicates a single position rather than a transition region. For example, it is unclear whether a cue point marks the beginning of a fade‑out or the exact moment playback should stop. Moreover, DJ transitions are context‑dependent, considering both the previous and next tracks. Yet, to our knowledge, no existing work explicitly models this relational aspect of mix point selection.

2.3 Music structure datasets

Various datasets have been developed for music structure analysis. A comprehensive overview is available in the work of Nieto et al. (2020), which reviews eight such datasets. Among these, the two most widely used are SALAMI (Smith et al., 2011) and the Harmonix Set (Nieto et al., 2019). SALAMI is the largest publicly available dataset for structure analysis, comprising 1,359 tracks across diverse genres such as classical, jazz, pop, world, and live recordings. It provides hierarchical segment annotations but lacks beat and downbeat information. Of the tracks, 1,048 (77%) were annotated by two independent annotators, while the remaining 335 (23%) were annotated by a single annotator. An annotator’s guide was provided, along with a controlled vocabulary for functional segment labels (e.g., intro, verse, chorus). The dataset has been used to study inter‑annotator agreement, highlighting the subjectivity involved in perceiving musical structure. The Harmonix set, in contrast, includes annotations for segments, beats, and downbeats, making it the largest dataset with all three types of annotations (912 tracks). This dataset primarily consists of Western popular music, including songs from the dance/electronic, hip hop, country, and rock genres. It was originally created to support rhythm‑action (beat‑matching) games. Each track is annotated by a single annotator. Although no formal annotation guide has been published, the vocabulary used for segment labels is diverse and often inconsistent, with many terms appearing only once or twice, suggesting the absence of a standardized label set.

In this study, we introduce the Raveform dataset, which includes beat, downbeat, and structural annotations specifically curated for EDM tracks. While several public datasets and models exist for music structure analysis, their applicability to EDM and DJ mixes remains limited. For example, despite the rhythmic regularity of EDM, existing beat and downbeat tracking models often exhibit degraded performance due to the lack of genre‑specific training data (see the cross‑dataset evaluation in Table 4). Furthermore, in analyses of Western popular music, song structure is often described in terms of verse–chorus forms, where the repetition and variation of lyrics play a central role in distinguishing sections (e.g., verses, choruses, bridges). This lyric‑centric perspective works well for much Western popular music but is less suitable for EDM, where tracks are frequently instrumental and structural perception is driven more by changes in rhythm, texture, and energy than by textual form. To address this gap, our annotation methodology adopts an energy‑centered perspective, focusing on perceived changes in musical intensity and function—an approach more aligned with how DJs interpret track structure during performance and mixing.

Our annotation methodology also diverges from those adopted in existing datasets. While independent annotations by multiple annotators are valuable for exploring the subjectivity of music structure perception, they often result in inconsistencies within the dataset. These inconsistencies can undermine the reliability and interpretability of structure analysis models trained on such data. For instance, annotators may disagree on whether a chorus segment begins with the vocal pickup before the downbeat or exactly at the downbeat. A model trained on conflicting annotations may adopt inconsistent boundary definitions, making its behavior difficult to predict or interpret. To address this issue, Raveform employs a collaborative annotation process in which each track is reviewed by multiple annotators, but the final output is a single, consolidated annotation. Annotators cross‑validated each other’s work and engaged in iterative discussions to resolve ambiguities and define consistent principles for segmentation and labeling. Through this process, we developed a domain‑specific vocabulary and a detailed set of annotation guidelines, ensuring both internal consistency and relevance to the DJing context.

3.1 Functions of segment labels

To complement the textual definitions below, we also provide audio examples for each segment function on the companion website.⁵

Intro. An intro appears at the beginning of a track and is specifically designed by producers to facilitate seamless mixing by DJs. Its primary function is to enable beat‑matching with the currently playing track. To serve this purpose, it typically consists of percussive elements with straightforward rhythmic patterns and minimal harmonic content, thereby avoiding potential key clashes. Common rhythmic components include four‑on‑the‑floor kicks (or bass drum hits) and/or backbeat snares. In EDM, which predominantly uses a $\frac{4}{4}$ time signature, ‘four‑on‑the‑floor’ refers to a pattern where a kick drum is played on every beat of a bar (beats 1, 2, 3, and 4), while the backbeat snare pattern places snare hits on beats 2 and 4. In terms of energy, the intro generally has higher energy than a breakdown but lower than that of a buildup or a drop. Its energy profile is typically steady or gradually increasing. Tension is usually low and stable throughout the segment, contributing to the intro’s unobtrusive character—an intentional design choice that allows it to blend smoothly with other tracks during a DJ mix.

Buildup. A buildup typically follows the intro and precedes either a breakdown or a drop. Its primary function is to progressively increase both energy and tension, although its energy level remains lower than that of the drop. This escalation is often achieved through the gradual addition of instruments. Not all tracks contain a distinct buildup; its presence depends on the producer's structural strategy. Tension can be developed within the buildup, the breakdown, or distributed across both, reflecting the producer’s approach.

Breakdown. A breakdown typically precedes a drop and serves as a crucial transitional segment characterized by a sharp reduction in energy and a progressive buildup of tension. It is often the segment with the lowest energy in the entire track. Initially, a breakdown may exhibit both low energy and low tension; however, as it unfolds, the tension intensifies, while energy remains minimal. This contrast amplifies the perceived impact of the forthcoming drop by creating a heightened sense of anticipation. To achieve this effect, producers frequently remove low‑frequency elements such as kick drums and basslines. Nonetheless, strategies for building tension in breakdowns vary widely. Some breakdowns consist solely of ambient textures, while others may incorporate sharp hi‑hats with complex rhythmic patterns.

Drop. The drop represents the core of an EDM track, where the primary musical idea is fully realized. Typically emerging after a breakdown or buildup, it marks the point of maximal energy and the release of accumulated tension. Drops are characterized by a dense sonic texture, often featuring the simultaneous presence of drums, bass, melodic elements, and various effects. Later drops within a track frequently exhibit greater energy than earlier ones, contributing to a sense of escalation across the track’s progression. Structurally, drops are usually followed by either a breakdown or a cooldown.

Cooldown. A cooldown segment follows a drop and precedes either a breakdown or an outro. Functionally, it serves as the inverse of a buildup, gradually reducing energy and tension. By providing a smoother transition, cooldowns help avoid abrupt shifts in intensity, particularly when moving toward structurally calmer sections. As with buildups, cooldowns are not universally present and are selectively employed by producers based on the desired flow and emotional contour of the track.

Bridge. As illustrated in Figure 9, a bridge segment appears infrequently in this dataset. When it does occur, it is typically situated near the middle of a track following a breakdown and preceding either another breakdown or a drop—and generally appears only once per track. Its role is analogous to that of a bridge in popular music: to introduce contrast, disrupt repetitive patterns, and re‑engage the listener’s attention. Musically, it often features distinct rhythmic, harmonic, or timbral elements that differentiate it from the surrounding sections.

Outro. An outro appears at the end of a track and shares many sonic characteristics with an intro, typically featuring percussive sounds and straightforward beats. Functionally, it serves the same purpose as an intro, facilitating smooth mixing for DJs. Unlike the intro, however, the energy in an outro usually remains steady or decreases slightly. Similar to the intro, the tension remains stable and low, ensuring that the segment blends seamlessly when transitioning to another track.

Ambient‑Intro. An ambient‑intro opens a track in a manner similar to an intro but is distinguished by the absence of percussive elements. Instead, it consists primarily of harmonic, melodic, and/or ambient sounds. This absence of rhythmic cues makes beat‑matching nearly impossible for DJs, requiring alternative mixing strategies. Although the term ambient‑intro is not as widely used as other functional segment labels, we introduce it here to capture this segment's distinctive sonic and functional characteristics.

Ambient‑outro. An ambient‑outro, located at the end of a track, is characterized by the absence of percussive elements. Like the ambient‑intro, this lack of rhythmic structure makes traditional DJ mixing techniques such as beat‑matching difficult. We introduce the term ambient‑outro—in parallel with ambient‑intro—to distinguish it from the more typical outro, which contains percussive elements.

3.2 Criteria for resolving ambiguity

Even though the functions of segment labels are described in the previous section, some ambiguity still remains in segmentation and labeling. This section outlines the criteria used to resolve such cases.

Regularity. EDM tracks often exhibit a regular phrase structure, with changes tending to occur at multiples of four or eight measures. Annotators used this regularity as a prior when placing boundaries: when several musically plausible options existed, they preferred boundaries that aligned with these phrase lengths. However, regularity was treated as a soft guideline rather than a hard constraint. When a clear musical change (e.g., in energy, rhythm, harmony, or instrumentation) occurred off the expected grid, boundaries were placed at the actual change, even if this resulted in sections of unusual length.

Consecutive segments with identical functions. Long passages with the same function—such as an extended drop—may span many measures; in our corpus, we even encountered drops lasting on the order of 80–90 measures. Such passages can either be treated as a single segment or split into multiple consecutive segments with the same label. We opt to divide these sections only when there is a clear musical change, such as a shift in energy, rhythm, chord progression, or instrumentation. Gradual evolutions are not used as dividing cues since identifying a precise boundary in such cases is inherently ambiguous. As a consequence, some long drops remain as a single segment, while others are split into several drops when discrete structural events occur. This policy reflects the balance between using regular phrase lengths as a guideline and maintaining consistent, musically motivated criteria across tracks. An example of a long drop that is split into three segments is provided as a listening example in the ‘Straightforward Example’ section of the companion website.

Drop‑like intro. Some tracks exhibit high energy right from the beginning, resembling a drop that lasts from the start to the middle until reaching a breakdown. In such cases, this section is classified as the intro, as it is likely the point where a DJ would perform beat‑matching.

Breakdown onset. In some tracks, the onset of a breakdown is ambiguous as the majority of instruments fade out gradually, ultimately reaching the lowest energy level. In such cases, the start is identified at the point where the energy begins to decrease, following the regularity rule previously described.

Cooldown and outro. When a track gradually removes instruments over a long period, it becomes difficult to distinguish between cooldown and outro segments. For example, the lead synth may fade out first, followed by the bassline after eight measures, and then the snares after another eight measures. While the changes in energy are noticeable, labeling becomes ambiguous. In such cases, we defined a segment as an outro when percussion elements dominate the texture. This allows for sequences like cooldown–cooldown–outro–outro, based on the sonic characteristics.

Excluded tracks. Out of 2,000 tracks, 577 were excluded for various reasons. This was primarily due to the noisy nature of crowdsourced data, as track links were annotated by users and audio files were collected from YouTube. Consequently, we discarded tracks that were not the original versions. This includes tracks with fading starts, short versions (or radio edits), and those created by clipping from a DJ mix. Additionally, we removed experimental EDM tracks that were intentionally made with ambiguous structures. Audio examples of such excluded tracks are provided in the ‘Ambiguous Example’ section of the companion website.

4.1 Overview

Table 3 summarizes key statistics of the dataset. The number of played tracks refers to the total instances of tracks played across all DJ mixes, including repeated appearances of the same track in different mixes. In contrast, the number of unique tracks counts each track only once, regardless of how often it appears. As a result, the number of played tracks exceeds the number of unique tracks. The number of transitions is smaller than the number of played tracks minus one because we only count a transition when two consecutive tracks in a mix are both successfully identified with available audio and (as detailed in Section 4.2) when their audio segments overlap after mix‑to‑track alignment. Out of 56,873 unique tracks, a subset of 1,423 tracks has been manually annotated with structural information by human experts. These annotated tracks represent the most frequently played selections across the mixes. Each annotation includes tempo, beat and downbeat positions, and functional segment boundaries with labels.

Figure 3 shows the genre distribution of the DJ mixes, while Figure 4 displays the genre distribution of the annotated tracks. Techno is the most represented genre in the dataset, followed by trance, drum and bass, and house, among others. Figure 5 presents the beats per minute (BPM) distribution of the annotated tracks. Tempo in our annotations follows a common EDM convention: when a stable four‑on‑the‑floor kick pattern is present, we set the tempo so that each kick corresponds to one beat. The most common BPM is 128, which aligns with typical tempos in genres such as house and techno. A secondary concentration appears in the 170–175 BPM range, primarily due to the presence of drum and bass tracks.

Figure 3

Figure 4

Figure 5

4.2 Data‑collection process

All data, except for the structural annotations, were sourced from MixesDB,⁶ a user‑contributed database for DJ mixes that provides metadata, audio links to the mixes, and links to the tracks played. To ensure accurate track identification, we used the site’s filtering feature. However, it is important to acknowledge that the MixesDB dataset contains inherent inaccuracies in its annotations. For instance, a track listed in the track list may not be the exact version played by the DJ. Additionally, dance music often includes various versions of a single track, such as short, long, or remix versions, which may not align with the version listed in the dataset. Figure 6 illustrates the distribution of the fraction of identified tracks per DJ mix; 342 mixes (7%) are fully identified.

Figure 6

To obtain time‑synchronized relations between mixes and tracks and to further filter out mismatched items, we estimate beat‑level mix‑to‑track alignments for all candidate mix–track pairs using the method of Kim et al. (2020). Each alignment specifies, at the beat level, which beat index in the track corresponds to which beat index in the mix and thereby yields estimated mix‑in and mix‑out positions for that track within the mix. We use the fraction of correctly aligned beats (match rate) as a quality measure and discard mix–track pairs whose match rate is lower than 0.1, treating them as misidentified or nonmatching tracks. Figure 7 shows the distribution of the match rate over all aligned track instances. For the dataset statistics reported in Table 3, we retain only alignments whose match rate is at least 0.1. The process of collecting the structural annotations will be detailed in the following subsection.

Figure 7

4.3 Structure‑annotation procedure

Music structure is subjective by its nature (Flexer and Grill, 2016; Nieto et al., 2020). As discussed in Section 2, our approach aims to incorporate perspectives from multiple annotators while establishing consistent annotation criteria. In essence, the goal is to make the inherently subjective judgments in structure annotation consistently applicable across different tracks. This subsection describes the annotation procedure and our efforts to standardize the annotation criteria.

Three experts participated in the annotation: two with DJing experience and one with a background in music production. The annotation process was organized into three phases. First, an annotator marked segment boundaries and assigned labels for each track. These initial annotations were then reviewed by a verifier. Finally, a final judge conducted a last round of validation. During the initial phase, annotators were encouraged to flag tracks with ambiguous structures. In the validation phases, reviewers marked any annotations they disagreed with. Weekly meetings were held throughout the process, during which annotators listened to ambiguous tracks together and refined the annotation criteria through discussion. The roles were rotated regularly among the annotators to ensure balanced participation.

The annotation process spanned 7 months. In the early stages, frequent changes to the segment label vocabulary necessitated multiple restarts. Over time, as the annotators collaboratively listened to and discussed ambiguous cases, disagreements became less frequent. Most disagreements were resolved through discussion, as annotators came to understand and appreciate each other’s perspectives. Often, disagreements arose from focusing on different musical elements—one annotator might focus on an added bassline, while another might pay more attention to the introduction of drums when identifying segment boundaries. In such cases, both perspectives were typically valid, and the choice came down to selecting the more consistent or useful interpretation. To guide such decisions, annotators considered questions such as: Which perspective is more helpful for DJs? Is this information necessary for DJing? Which view better reflects energy‑centric transitions? Can this criterion be applied consistently across other tracks?

In many cases, the challenge was not disagreement but rather the intrinsic ambiguity of a track’s structure. For example, in EDM, particularly in repetitive and hypnotic techno, musical elements often fade in and out gradually. Although one can clearly identify a drop and an outro, it may be difficult to pinpoint the exact boundary between them due to the smooth, continuous decrease in energy and loudness. Capturing such ambiguity in a consistent and descriptive manner is inherently difficult. Therefore, collaborative listening and discussion were essential components of the annotation process.

Annotations were created using Rekordbox,⁷ a DJ software platform. Segment boundaries and labels were recorded using the memory cue feature, while beats and downbeats were annotated with the help of Rekordbox’s beat and downbeat tracking feature (‘track analysis’). In our experience, this analysis works reliably for tracks with approximately constant tempo and four‑on‑the‑floor drum patterns, which covers the majority of our corpus, although downbeat positions can occasionally be offset. For tracks with substantial tempo changes or more irregular rhythms, the estimated beat grid can become unstable. For every annotated track, we inspected the beat grid, corrected local beat/downbeat errors when necessary, and excluded tracks whose automatic estimates were too inaccurate to be corrected with reasonable effort. Annotations were then exported as XML files and processed using Python. Track genres were assigned by referencing the genre tags listed for each track on Beatport.⁸

5.1 Segment‑level statistics

Figure 8 illustrates the distribution of the number of segments per track. Most tracks contain between 8 and 13 segments, a range that closely aligns with the segment distribution observed in the Harmonix set (Nieto et al., 2019). Figure 9 shows the distribution of segment labels across the dataset. Given that the dataset includes 1,423 tracks, labels such as ‘Ambient‑Intro,’ ‘Ambient‑Outro,’ and ‘Bridge’ appear infrequently, occurring far less often than the total number of tracks. In contrast, segments such as ‘Intro,’ ‘Outro,’ ‘Buildup,’ and ‘Cooldown’ typically appear once per track, while ‘Drop’ and ‘Breakdown’ tend to occur multiple times within a single track. Finally, Figure 10 presents the distribution of segment lengths, with the top portion showing lengths in seconds and the bottom portion showing such in measures. The most common segment length is 30 seconds or 16 measures. This peak corresponds to tracks with the most prevalent tempo of 128 BPM, where a 16‑measure segment typically spans 30 seconds.

Figure 8

Figure 9

Figure 10

5.2 Quantitative analysis

Figure 11 illustrates the common structure of EDM tracks, acquired by aggregating the structural annotations. To analyze common structural patterns, we first normalized the lengths of all tracks to a standard unit length. Then, at each point along this normalized timeline, we calculated the proportion of tracks assigned to each segment label. This approach allowed us to visualize how often each segment label occurs at each point in a typical EDM track. The visualization reveals a typical structure in EDM tracks: intro–buildup–drop–breakdown–drop–cooldown–outro.

Figure 11

Figure 12 illustrates the average loudness (dB) values across frequency bands for each segment label. As we defined the energy characteristics of the labels, the quantitative measure of energy, represented by dB, aligns with this definition. The drop possesses the highest energy, while segments such as the ambient‑intro and ambient‑outro, which lack many percussive instruments, exhibit the lowest energy. Similarly, the breakdown segment, characterized by the absence of many instruments, also displays low energy. Additionally, the energy increases progressively from intro to buildup to drop and conversely decreases from drop through cooldown to outro.

Figure 12

6.1 Experimental setup

6.2 Results

Table 4 summarizes the cross‑dataset evaluation results of metrical and functional structure analysis. The top four rows correspond to the test set from the Harmonix set, while the bottom four rows pertain to the Raveform test set. For each test set, the first row reports the performance of the model trained on the training split of the same dataset. The second row shows the model trained on the cross‑dataset training split; the third row corresponds to the model trained using both datasets; and the fourth row presents the results of Madmom, which has a different architecture and was trained on 10 datasets, excluding both the Harmonix set and Raveform.

Overall, the performance on Raveform is higher than that on the Harmonix set, suggesting that Raveform provides more straightforward learning targets. In particular, scores for segmentation and labeling are consistently higher under the same experimental conditions, except for the dataset. Given the subjective nature of segment boundaries and labels, this suggests that Raveform’s annotations are more consistent. As expected in cross‑dataset evaluation, performance degrades when models are tested on a dataset different from the one they were trained on. Notably, the performance drop is more pronounced for downbeats, segments, and segment labels that involve more abstraction compared to beats. Models trained on both datasets exhibit comparable performance to models trained and evaluated on the same dataset. Madmom shows competitive results in beat tracking but lags significantly in downbeat tracking.