Osu2MIR: Community-Driven Beat Annotation

• Osu2MIR is a dataset-construction pipeline that converts Osu! beatmaps into beat and downbeat annotations, offering a scalable alternative to traditional expert labeling.
• It employs a systematic extraction process that parses chart files, applies a 5-second heuristic to partition timing points, and ensures reliable timing structure.
• The resource targets underrepresented genres like anime, Vocaloid, and video game music, leveraging community consensus to mitigate annotation biases.

Searching arXiv for the named paper and closely related MIR beat-tracking references. Osu2MIR is a dataset-construction pipeline and released resource that converts Osu! beatmaps into beat and downbeat annotations for music information retrieval (MIR). It is introduced as a scalable alternative to traditional expert-annotated beat-tracking datasets, with the central premise that Osu!, a community-authored rhythm game ecosystem, already contains timing structure suitable for MIR supervision. The resource is motivated by limitations of existing beat and downbeat datasets, including small scale, narrow genre coverage, slow annotation throughput, and the biases of single annotators. Osu2MIR is positioned as particularly relevant for underrepresented repertoires such as anime, Vocaloid, and video game music, and as evidence that community-produced rhythm-game artifacts can encode musically meaningful temporal structure (Liu et al., 16 Sep 2025).

1. Source Domain and Research Motivation

Osu2MIR is built on the observation that Osu! beatmaps are created and refined by a large, diverse community, and that many songs receive multiple independent annotations. The source platform is described as hosting over 40,000 beatmap sets and being continuously updated by a global community. This makes it potentially useful not only as a source of more data, but also as a source of more varied data and multiple annotations per song, which may help mitigate individual annotator bias (Liu et al., 16 Sep 2025).

The resource addresses a practical bottleneck in MIR, especially for beat tracking and downbeat annotation. Deep learning systems for beat tracking depend on supervised data, but expert-created labels are expensive and difficult to scale. Existing datasets are characterized as lagging behind current musical trends and often omitting fan and game-music communities that are prominent in rhythm-game ecosystems. Within that framing, Osu! is attractive because it already contains timing structure created for synchronization between music and game interaction.

A notable aspect of the motivation is that the project does not treat rhythm-game charts as arbitrary gameplay artifacts. Instead, it treats ranked beatmaps as socially filtered temporal annotations. This suggests that the ranked ecosystem functions as a community-driven quality-control mechanism, although the paper does not formalize that mechanism statistically. The main empirical support for this interpretation is the reported high consistency across multiple user-created beatmaps of the same song (Liu et al., 16 Sep 2025).

2. Data Acquisition and Beatmap Selection

The data source is Osu! beatmap sets downloaded using Batch Beatmap Downloader. Collection is restricted by four explicit filters:

1. only ranked beatmap sets
2. only Osu! mode beatmaps
3. only beatmap sets containing an “Insane” difficulty
4. only beatmap sets with at least 200 favorites

The rationale for these filters is procedural rather than statistical. Ranked beatmaps have passed the platform’s ranking process and therefore are expected to meet quality standards. Osu! mode is the most popular mode and is the mode targeted in the study. Requiring an “Insane” difficulty is used as a proxy for sufficient rhythmic complexity, while requiring at least 200 favorites acts as a quality and popularity filter (Liu et al., 16 Sep 2025).

These criteria shape the resulting corpus in technically important ways. They privilege socially validated charts and exclude less established or less visible material. A plausible implication is that Osu2MIR does not represent the full Osu! ecosystem uniformly, but rather a curated subset aligned with annotation reliability and practical utility for beat-tracking research. The paper does not provide a full genre breakdown, file count by genre, audio duration totals, or a precise license statement, so corpus composition can only be characterized at the level explicitly reported.

The study also does not claim that the extracted resource is a direct substitute for all traditional MIR datasets. Rather, it presents Osu! as an alternative annotation source with distinct advantages in scale potential, diversity, and updateability. The emphasis is on using an existing community corpus to alleviate annotation scarcity, not on demonstrating superiority through a head-to-head training benchmark (Liu et al., 16 Sep 2025).

3. Extraction Pipeline and Annotation Semantics

The extraction pipeline operates on .osz beatmap archives. Each archive is decompressed to recover .osu chart files and the associated audio, typically .mp3. The key file sections are [General], which identifies the referenced audio filename, and [TimingPoints], which contains the timing structure used for beat and downbeat derivation. The method distinguishes uninherited timing points from inherited timing points, and only the uninherited timing points are used because they encode tempo and meter structure (Liu et al., 16 Sep 2025).

The procedural workflow is explicit. For each beatmap, the archive is decompressed, the .osu metadata is parsed, the audio file named in [General] is identified, and the uninherited timing points in [TimingPoints] are extracted. For each uninherited timing point, the beat length and meter are read, that point is treated as the start of a timing region, and the end of the region is defined by the next uninherited timing point or by the end of the song if no subsequent point exists. Beat timestamps are then generated by iteratively adding the beat length from the region start until the region end, and beat indices within the bar are assigned using the meter. The annotations are saved as tab-separated text files, and the corresponding audio file is copied into the target dataset directory (Liu et al., 16 Sep 2025).

No formal mathematical equations are given for the extraction process. The method is therefore best understood as a deterministic parsing-and-stepping procedure rather than as an optimization or estimation framework. The released annotations are described as metered beat annotations. This suggests that downbeats are inferred from metrical position: if beat indices are positions within the bar, then the first beat in each measure is the downbeat. The paper does not specify a symbolic mapping such as “index 1 = downbeat,” but that is the implied mechanism when it states that the released files contain metered beat labels (Liu et al., 16 Sep 2025).

The annotation format is narrowly specified. Audio files are copied from beatmap archives, generally as .mp3, and annotations are stored in tab-separated text files. No alternative representation such as JAMS, CSV, or MIDI is mentioned. The repository associated with the project is https://github.com/ziyunliu4444/osu2mir, and it is described as containing customizable pipelines for extracting annotations from Osu! beatmaps into beat/downbeat-style files (Liu et al., 16 Sep 2025).

4. Timing-Point Partitioning as a Curation Heuristic

A central contribution of Osu2MIR is its partitioning of beatmaps into three subsets according to the spacing between uninherited timing points:

1. beatmaps with a single uninherited timing point
2. beatmaps with multiple uninherited timing points separated by at least 5 seconds
3. beatmaps with multiple uninherited timing points where at least some adjacent points are separated by less than 5 seconds

The threshold is exactly 5 seconds and is described as a heuristic chosen to distinguish genuinely distinct timing sections from minor tempo fluctuations within a section. This partition is the main curation device in the work, because the authors argue that timing-point structure strongly predicts annotation reliability (Liu et al., 16 Sep 2025).

The single-timing-point subset is interpreted as rhythmically simple and mostly reliable. Manual checks reportedly confirm high quality even in syncopated examples and in non-4/4 examples. Multiple annotators are also reported to show strong agreement on these songs. The limitation of this subset is not annotation quality but limited rhythmic diversity, which may reduce its value for training models that require broader structural variety.

The subset with multiple timing points separated by at least 5 seconds is presented as the most valuable group. It contains more complex rhythms and deliberate meter changes, yet the annotations generally align well with the music. Manual review suggests that annotators often agree on tempo and meter, although some interpretation differences occur, such as tempo-halving variants. This subset is both high-quality and more diverse than the single-timing-point subset, and it is the source of the released high-quality dataset osu2beat2025 (Liu et al., 16 Sep 2025).

The subset with adjacent timing points closer than 5 seconds is explicitly flagged as requiring extra filtering and caution. The paper attributes the noisier behavior of this group to several causes: very high annotation density caused by expressive timing, wrong meter settings, limitations in the Osu! editor, specifically no support for non-4-based meters, and inconsistent annotation goals such as syncing to lyrics instead of musical beat structure. In this view, dense timing-point placement often reflects local charting hacks rather than stable musically meaningful tempo and meter annotation.

This three-way partition is therefore more than a descriptive taxonomy. It functions as a practical curation heuristic with direct implications for dataset construction. Widely spaced timing changes are treated as more likely to correspond to real structural or metrical changes in the music, whereas densely packed changes are treated as more likely to encode annotation artifacts or author-specific gameplay intentions (Liu et al., 16 Sep 2025).

5. Reliability Analysis and Released Subset

Quality analysis combines automatic comparison with manual inspection. The automatic stage runs the RNN beat and downbeat tracker from madmom on the corresponding audio, and compares the predictions to the Osu!-derived labels using standard beat-tracking metrics: F-measure, CMLt, and AMLt, for both beats and downbeats. These metrics are used only as heuristic indicators of unusual or questionable cases, not as definitive proof of ground-truth quality, because low agreement could arise either from poor Osu! annotation or from model failure, especially on unusual music. Manual analysis then reviews diverse examples with low or inconsistent madmom agreement, compares multiple user annotations for the same audio, and performs targeted inspection informed by the characteristics of each subset (Liu et al., 16 Sep 2025).

The reported mean metric values for the three subsets are as follows:

Subset	Beat metrics	Downbeat metrics
Single timing point	F1 0.91, CMLt 0.82, AMLt 0.92	F1 0.87, CMLt 0.80, AMLt 0.91
Multiple timing points, all adjacent changes at least 5 s apart	F1 0.89, CMLt 0.77, AMLt 0.87	F1 0.81, CMLt 0.72, AMLt 0.82
Multiple timing points with some adjacent changes under 5 s	F1 0.86, CMLt 0.71, AMLt 0.83	F1 0.76, CMLt 0.61, AMLt 0.73

These values are interpreted qualitatively rather than as a formal benchmark. The single-timing-point subset is described as reliable but rhythmically simple. The >=5 s multi-timing-point subset is characterized as high-quality and more diverse, which is why it forms the basis of the released resource. The \<5 s subset is treated as the least trustworthy and as needing additional curation (Liu et al., 16 Sep 2025).

The released dataset subset is osu2beat2025. It is drawn from the second partition, namely beatmaps with multiple uninherited timing points spaced at least 5 seconds apart. The release statistics are 741 annotations and 708 distinct audios, indicating that some audio tracks have multiple annotation versions. The dataset contains metered beat annotations, and the filename format is explicitly MD5_BeatmapSetID_beats_metered.txt (Liu et al., 16 Sep 2025).

Two practical caveats are emphasized. First, some labels may extend beyond the audible range, so trimming the audio to the effective end may be advisable when building a custom dataset or evaluation setup. Second, when duplicate versions of the same song contain conflicting meter annotations, the version using a compound or additive meter is “typically the correct interpretation.” This is one of the few explicit curation recommendations in the release description and is technically important when selecting among multiple annotations of the same audio (Liu et al., 16 Sep 2025).

6. Community Agreement, Limitations, and Research Outlook

One of the most consequential findings is the reported consistency across multiple user annotations of the same song. The study selects audio tracks with multiple beatmaps authored by different users and compares them manually. The conclusion is positive: multiple annotators show strong agreement in the single timing-point subset, and annotators in the >=5 s multiple-timing-point subset often agree on tempo and meter. When disagreements arise, they are often interpretable, for example as tempo-halving variants rather than arbitrary mismatches (Liu et al., 16 Sep 2025).

This supports the view that Osu! can function as a community-driven annotation source rather than merely as a repository of idiosyncratic gameplay charts. If independently authored beatmaps converge on similar beat structures, then the community is implicitly encoding musically meaningful temporal structure. At the same time, the paper is explicit about what it does not provide: there is no formal inter-annotator agreement coefficient such as Cohen’s kappa, no exact duplicate-song sample sizes for the agreement analysis, and no track-level percentages of manually accepted or rejected examples. The agreement claims are therefore qualitative rather than statistically parameterized.

Several limitations and biases are also identified. The subset with timing points closer than 5 seconds often needs extra curation. Osu! editor limitations, especially the lack of support for non-4-based meters, can distort annotations. Some chart authors may synchronize to lyrics or other salient events instead of musical beats. Labels may extend beyond the audible audio. The single timing-point subset may be too rhythmically simple to maximally benefit model training. More broadly, because the data is sourced from a game community, charting conventions may differ from traditional MIR annotation conventions (Liu et al., 16 Sep 2025).

The analysis includes an additional observational finding: user annotations tend to occur slightly ahead of beat onsets predicted by madmom. The authors suggest that this may reflect human perception or internalization of rhythm. This is not formalized into a quantitative offset analysis, but it indicates that community timing labels may encode perceptual beat placement rather than purely onset-locked timing.

Future directions are framed in terms of broader reuse of Osu! data beyond beat tracking. The authors suggest building tools that help beatmap creators, developing pipelines for other MIR tasks using Osu! data, fostering an open-source ecosystem around these resources, and exploring language- and vocal-style-specific questions such as Japanese/Korean lyric alignment and Vocaloid timing behavior. Within that outlook, Osu2MIR is best understood not only as a dataset release but also as a methodological proposal: that community-authored rhythm-game data can be systematically transformed into MIR-ready supervision when extraction is coupled with timing-structure-aware curation (Liu et al., 16 Sep 2025).