---
title: 'osu2beat2025: Multi-Domain Beat Annotation & Systems'
url: https://www.emergentmind.com/topics/osu2beat2025
type: topic
---

# osu2beat2025: Multi-Domain Beat Annotation & Systems

Across recent arXiv literature, **osu2beat2025** does not denote a single canonical artifact. Instead, it appears as a high-quality **beat and downbeat annotation subset derived from Osu! beatmaps**, as a **target name for implementation-oriented beat tracking, melody generation, and chart generation pipelines** built around osu!-style rhythmic structure, and as the **system identifier attached to Oregon State University’s BeaverTalk submission** to the IWSLT 2025 simultaneous speech translation task [2509.12667] [2407.21658] [2306.06284] [2311.13687] [2505.24016]. The term is therefore best understood as a context-dependent label whose meaning is fixed by the surrounding paper: in music information retrieval it is tied to Osu!-derived rhythmic data and beat-synchronous generation, whereas in speech translation it labels an unrelated cascaded streaming system.

## 1. Terminological scope and disambiguation

The literature associates the same string, **osu2beat2025**, with several distinct research objects. The most concrete and dataset-like usage is in Osu2MIR, where it refers to a released subset of metered beat annotations derived from Osu! beatmaps. Other papers use the name as a target application label for rhythm-analysis or generation pipelines built around osu!-style timing structure. A separate, unrelated usage appears in simultaneous speech translation, where **BeaverTalk (osu2beat2025)** is the identifier of Oregon State University’s IWSLT 2025 system [2509.12667] [2505.24016].

| Usage | Domain | Source |
|---|---|---|
| High-quality metered beat/downbeat annotation subset derived from Osu! beatmaps | MIR dataset curation | [2509.12667] |
| DBN-free beat/downbeat tracking pipeline aligned to a pipeline like osu2beat2025 | Beat tracking | [2407.21658] |
| Beat-to-melody generation system taking osu!-style rhythmic inputs | Symbolic music generation | [2306.06284] |
| Beat-aligned spectrogram-to-sequence system for rhythm-game chart generation | Chart generation | [2311.13687] |
| BeaverTalk (osu2beat2025), Oregon State University’s cascaded streaming speech translation system | Simultaneous speech translation | [2505.24016] |

A common misconception is to treat osu2beat2025 as a single benchmark or model family. The published record instead shows an overloaded label spanning dataset release, pipeline adaptation, and an unrelated team/system name. This suggests that the term functions as a project-local designation rather than a standardized community-wide ontology.

## 2. Osu2MIR subset: beat and downbeat annotations from Osu! timing metadata

In **"Osu2MIR: Beat Tracking Dataset Derived From Osu! Data"** [2509.12667], osu2beat2025 is presented as a **high-quality, metered beat and downbeat annotation subset derived from Osu! beatmaps**. The project motivation is explicit: Osu! offers a large, diverse, community-generated source of timing annotations, with over **40,000 beatmap sets** and strong representation of **anime, Vocaloid, and video game music**, thereby complementing traditional MIR datasets that are small, slow to grow, and genre-limited.

The extraction pipeline operates on `.osz` archives, which are decompressed to obtain `.osu` files and the referenced audio. The `[General]` section identifies the audio file, and the `[TimingPoints]` section provides both inherited and uninherited timing points. Only **uninherited timing points** are used for beat and downbeat generation. For each uninherited timing point with onset time $t_0$, beat length in milliseconds, and meter, the script generates beats until the next uninherited timing point or the end of the audio. The core timing equations are stated as

$$
\mathrm{BPM} = \frac{60000}{\mathrm{beatLength}_{ms}}
$$

and

$$
t_n = t_0 + n \cdot \frac{60}{\mathrm{BPM}}.
$$

The resulting annotation files store **beat times in seconds** and **beat indices within the bar**.

The dataset is partitioned by the spacing of uninherited timing points into three subsets: **single timing point**, **multiple uninherited timing points spaced at least 5 seconds apart**, and **multiple uninherited timing points spaced less than 5 seconds apart**. osu2beat2025 corresponds to the second category, namely beatmaps with **multiple tempo or meter sections separated by at least 5 seconds**. The paper describes the **5-second threshold** as a practical heuristic for distinguishing deliberate section changes from minor timing fluctuations. Beatmaps with more closely spaced timing points are excluded because they often encode expressive micro-adjustments, editor limitations, or non-beat-synchronous goals such as lyric alignment.

The reported release statistics state that **741 annotations for 708 distinct audio tracks** are released across the project’s high-quality subset, and the paper attributes the public high-quality subset to this partition in the release. The benchmark alignment numbers for the subset forming the basis of osu2beat2025 are reported against **madmom’s RNN beat and downbeat tracker** as follows [2509.12667]:

| Metric group | Beat | Downbeat |
|---|---:|---:|
| F1 | 0.89 | 0.81 |
| CMLt | 0.77 | 0.72 |
| AMLt | 0.87 | 0.82 |

The file format is a **tab-separated text file** named `MD5_BeatmapSetID_beats_metered.txt`, with each row containing beat time in seconds and beat index within the bar. The paper also notes a practical caveat: some beat labels can extend beyond the audible range because of trailing timing definitions, so downstream dataset construction may require trimming to the effective audio end.

This usage gives osu2beat2025 its most dataset-like meaning. In that sense, it is a curated Osu!-derived annotation resource designed for training and evaluating beat and downbeat trackers under tempo and meter changes that are section-level rather than micro-timed.

## 3. DBN-free beat and downbeat tracking aligned to osu2beat2025 workflows

In the implementation-oriented guide derived from **"Beat this! Accurate beat tracking without DBN postprocessing"** [2407.21658], osu2beat2025 is treated as the target environment for a **general-purpose beat and downbeat tracker without DBN postprocessing**. The paper’s central argument is that Dynamic Bayesian Network postprocessing enforces tempo and meter constraints that can fail on diverse music, especially on time-signature changes, rubato, concatenated tracks, and material outside fixed tempo ranges. Removing the DBN is therefore presented as a route to greater generality.

The input representation is specified as **30 seconds of mono audio resampled to 22.05 kHz**, converted into a **128-bin mel spectrogram from 30 Hz to 10 kHz**. The STFT parameters are a **window size of 1024 samples** and **hop size of 441 samples**, yielding approximately **50 frames per second**. Magnitudes are scaled as

$$
\ln(1 + 1000x),
$$

and an initial batch normalization is applied across frequency bands.

The network architecture alternates **convolutions** with **partial transformers** directed over frequency and time axes, followed by a **full temporal transformer**. The frontend begins with a 2D convolution of kernel size **$3 \times 4$** and three repeated blocks, each consisting of a frequency-directed partial transformer, a time-directed partial transformer, and a frequency-strided convolution that halves the number of bands and doubles the channel count. After reshaping from $T \times 4 \times 256$ to $T \times 1024$, the model projects to $T \times 512$ and applies **6 stacked transformer blocks** with **16 attention heads** and **head size 32**. The output layer has separate beat and downbeat heads, with a **Sum Head** that sets beat logits to beat-linear-output plus downbeat-linear-output, encouraging downbeats to coincide with beats without making that constraint hard.

The training objective is a **shift-tolerant weighted BCE**. The guide states the loss as

$$
L_{st}(y, \hat{y}, w) = -\sum_t w y_t \log(m_7(\hat{y})_t) + (1 - m_{13}(y)_t)\log(1 - m_7(\hat{y})_t),
$$

where predictions are max-pooled over a **7-frame window** before comparison, and negative contributions near positives are ignored through a **13-frame pooling mask** on the labels. This design is intended to align the training objective with the **$\pm 70$ ms** evaluation tolerance while producing sharp local peaks.

The training regimen spans **18 datasets**, including Simac, SMC, Hainsworth, Ballroom, HJDB, Beatles, Harmonix, multiple RWC subsets, TapCorrect, JAAH, Filosax, ASAP, Groove MIDI, GuitarSet, and Candombe, with **GTZAN** reserved as **test-only**. Optimization uses **AdamW**, **100 epochs**, **batch size 8**, **gradient accumulation over 8 batches**, a learning-rate warm-up to **0.0008** for **1000 steps**, and cosine annealing. The guide reports training times of approximately **8 h on an RTX 2080 Ti**, **6 h on an A40**, and **4 h on an A100**.

Inference is deliberately minimal. Beat and downbeat probabilities are thresholded at **0.5**, peaks are selected by local non-maximum suppression within a **$\pm 3$-frame neighborhood**, and downbeats are snapped to the nearest beat. On **GTZAN**, the DBN-free system is reported at **Beat F1 89.1 ± 0.3**, **CMLt 79.8 ± 0.6**, **AMLt 89.8 ± 0.4**, and **Downbeat F1 78.3 ± 0.4**, **CMLt 67.3 ± 0.8**, **AMLt 79.1 ± 0.6**; the DBN comparison improves continuity but lowers F1. The paper therefore distinguishes between **event accuracy** and **continuity**, a distinction that matters for downstream rhythm-game or beatmap workflows.

Within the osu2beat2025 context, this line of work defines a DBN-free, high-resolution, beat/downbeat detection stack whose main appeal is broad metrical coverage rather than strict periodic smoothing.

## 4. Beat-conditioned melody generation from osu!-style rhythmic inputs

In the adaptation of **"Everybody Compose: Deep Beats To Music"** [2306.06284], osu2beat2025 is described as a **beat-to-melody generation system** that takes **osu!-style rhythmic inputs** and produces **monophonic melodies aligned with those beats**. The formulation is explicitly conditional: given a beat sequence $x = (x_1, \ldots, x_T)$, the model generates a pitch sequence $y = (y_1, \ldots, y_T)$ under the autoregressive factorization

$$
p(y_t \mid y_{<t}, x_{\le t}).
$$

This is contrasted with unconditional note-sequence modeling $p(y_t \mid y_{<t})$.

The core beat representation is a per-step tuple **$(\mathrm{rest}, \mathrm{dur})$**, where $\mathrm{rest}$ is the time after the release of the previous note and $\mathrm{dur}$ is the duration of the current note. Labels are pitches in **MIDI 0–127**, with **0 reserved for start-of-sequence**. For osu!-style inputs, the adaptation maps **TimingPoints** and **HitObjects** into this representation: circles become short-duration hits, sliders and spinners become sustained segments, and additional information such as **hitsound class**, **slider velocity multiplier**, **inter-onset interval**, **bar accent**, and **section ID** may be concatenated as conditioning features.

Three model families are proposed. The first is **LSTM with Full Attention**, using a bidirectional LSTM encoder and an attention-equipped decoder. The second is **LSTM with Local Attention**, in which the decoder uses the encoder state at the same timestep rather than attending globally; the guide states that this variant excelled in validation accuracy and qualitative results for the original task. The third is a **Transformer with Relative Position Representation**, where relative-position terms explicitly encode rhythmic distances and support phrase-level structure. The relative attention is written as

$$
\mathrm{RelativeAttention}(Q, K, V) = \mathrm{Softmax}\left(\frac{QK^T + QE^{rT}}{\sqrt{d_k}}\right)V.
$$

Training uses the autoregressive negative log-likelihood

$$
L = - \sum_t \log p(y_t \mid y_{<t}, x_{\le t}),
$$

with teacher forcing. The adaptation recommends transposition, tempo scaling, and meter shifts, and it proposes sequence slicing to fixed lengths such as **256–512** timesteps per epoch. Suggested defaults include **Adam or AdamW**, learning rate **$10^{-3}$**, **gradient clipping at 1.0**, **batch size 32–64**, and embedding or hidden dimensions in the ranges given in the guide.

The base dataset in the original project is **MAESTRO MIDI**, preprocessed into monophonic melody lines using note-seq melody inference. For osu2beat2025, however, the guide centers the conversion of `.osu` files into beat tokens: parse timing points, compute object start and end times, derive rest and duration, quantize to a **PPQ** grid such as **PPQ=480**, and then generate or export MIDI from predicted pitches. Reported objective metrics include token accuracy, with **LSTM local attention** listed at approximately **0.5077** and **Transformer RPR** at approximately **0.4690**.

This use of osu2beat2025 places the term in the symbolic-music domain: not as a released dataset, but as a beat-conditioned melody generation framework grounded in osu!-style rhythm annotations and interactive workflows such as **tapping** and **recoloring**.

## 5. Beat-aligned spectrogram-to-sequence chart generation

In the adaptation of **"Beat-Aligned Spectrogram-to-Sequence Generation of Rhythm-Game Charts"** [2311.13687], osu2beat2025 denotes a **beat-aligned spectrogram-to-sequence system for automatic rhythm-game chart generation**. The formulation departs from framewise event detection and instead casts chart generation as **conditional sequence generation** over beat-synchronous acoustic features.

The chart vocabulary is compact. Each chart event is represented by two tokens: a **time token** in the range **0–95**, indexing a **2-beat region subdivided into 96 equal grid steps**, and an **action token** in the range **97–176**, indexing one of **80 permissible player actions**. The vocabulary also includes a **separator token** and an **EOS token**. This tokenization yields a discrete event sequence rather than a heavily imbalanced framewise label space.

The acoustic frontend is beat synchronous. Audio is transformed to mono and converted into **log-Mel spectrograms** with **80 Mel bands**, **FFT size $N = 512$**, **window size $W = 512$**, and a **hop size corresponding to $1/48$ of a beat at the current tempo**. Over a song of $B$ beats, the beat-synchronous spectrogram has **$48B$ frames**, and each **4-beat** encoder window spans **192 time steps**. The alignment equations are given as

$$
t_k = k \cdot \frac{60}{\mathrm{BPM}\cdot s}, \qquad s=48,
$$

for constant tempo, and

$$
A(t) = \arg\min_k |t - t_k|.
$$

The model is a **3-layer encoder-decoder Transformer** with **model input dimension 256**. The encoder consumes four beats of beat-aligned log-Mel features, and the decoder is conditioned on the **last seven chart tokens from the previous two beats** plus a **48-dimensional difficulty embedding**. Generation proceeds **two beats at a time** in an autoregressive sliding-window fashion.

Training uses **cross-entropy with label smoothing 0.02**, **Adam**, learning rate **$2\times10^{-4}$**, **batch size 32**, and **10 epochs**. Pretraining on the large **osu!mania 4-key dataset** is central to the reported performance, and finetuning on **Fraxtil** and **ITG** is carried out for **4 epochs** at learning rate **$2\times10^{-5}$**. The curated osu!mania dataset sizes are reported as **1,598 songs / 5,494 charts / 2.24M beats** for training, **199 songs / 634 charts / 252k beats** for validation, and **207 songs / 643 charts / 255k beats** for test.

Evaluation uses **micro-F1** on event times. On osu!mania, the proposed model achieves **84.6%**, compared with **83.0** for CNN and **83.9** for C-LSTM. Pretraining improves transfer performance to **Fraxtil** and **ITG**, reaching **78.9%** and **74.3%**, respectively. The beat-group ablation is particularly important: aligned training dramatically improves uncommon subdivisions, with **12th-note F1 59.2** versus **9.2** for unaligned training, and **24th-note F1 26.3** versus **3.94**.

In this setting, osu2beat2025 refers to a beat-aligned, audio-to-chart pipeline for rhythm-game content generation. Its distinctive assumption is that temporal alignment to the beat grid is not merely a preprocessing convenience but an integral modeling choice.

## 6. BeaverTalk and the non-MIR naming collision

The speech-translation paper **"BeaverTalk: Oregon State University's IWSLT 2025 Simultaneous Speech Translation System"** [2505.24016] uses the identifier **BeaverTalk (osu2beat2025)** for a completely different system: a **cascaded, streaming speech-to-text translation system** for the **IWSLT 2025 simultaneous translation task**. This system targets the **text-to-text track**, where unsegmented audio is first transcribed and then translated under latency constraints.

The architecture is a left-to-right cascade consisting of **Silero VAD**, **Whisper Large V2** for ASR, and **Gemma 3 12B** for simultaneous translation. The VAD segmentation is controlled by **maximum segment duration (MSD)**, **voice probability threshold (VPT)**, and **maximum unvoiced duration (MUD)**; the paper reports that **MUD > 0.1 s degrades performance**, and fixes **MUD = 0.1 s**. Whisper uses a **stable transcription policy** that commits only the repeated prefix across consecutive intervals, together with a transcript context buffer. The translation model is **fine-tuned with LoRA adapters** under quantization, with **rank $r=64$**, **scaling $\alpha=16$**, and **dropout $0.1$**, using a conversational prompting strategy with a **single prior-sentence memory bank in the source language**.

The task directions are **English$\rightarrow$German** and **English$\rightarrow$Chinese**, each evaluated under low- and high-latency bands. On the **ACL 60/60 development set**, the reported results for the final **12B** system are: **English$\rightarrow$German BLEU 24.64 at StreamLAAL 1837.86** in low latency and **BLEU 27.83 at StreamLAAL 3343.73** in high latency; **English$\rightarrow$Chinese BLEU 34.07 at StreamLAAL 2216.99** in low latency and **BLEU 37.23 at StreamLAAL 3521.35** in high latency. Training and evaluation of the **Gemma 3 12B** system were conducted on an **NVIDIA H200**.

This usage is unrelated to Osu!, beat tracking, beatmaps, or music generation. Its inclusion matters because it demonstrates that **osu2beat2025** is not a uniquely music-specific term in the 2025 arXiv record. A second common misconception is therefore that every occurrence of the label refers to Osu!-derived MIR resources. The BeaverTalk paper shows that this is not the case.

## 7. Conceptual unities and limits across usages

Despite the naming collision, the music-oriented usages of osu2beat2025 exhibit a coherent technical center. They all treat **rhythmic structure** as the organizing variable: Osu2MIR extracts beat and bar-position annotations from timing metadata; the DBN-free beat tracker estimates beats and downbeats directly from audio without tempo- or meter-hardening postprocessing; the beat-to-melody system conditions symbolic pitch generation on rhythmic tuples such as $(\mathrm{rest}, \mathrm{dur})$; and the beat-aligned chart generator transforms audio into beat-synchronous event sequences [2509.12667] [2407.21658] [2306.06284] [2311.13687].

Another shared property is **explicit attention to tempo and meter changes**. In Osu2MIR, the 5-second partition distinguishes section-level tempo and meter changes from dense micro-adjustments. In the DBN-free tracker, removal of DBN priors is motivated by failure modes under variable tempo and meter. In beat-conditioned melody generation, TimingPoints and section IDs enter the conditioning stream. In beat-aligned chart generation, the hop size itself is defined as **$1/48$ of a beat at the current tempo**. These convergences suggest that osu2beat2025, in its music-related senses, belongs to a broader methodological shift toward **beat-synchronous, section-aware modeling** rather than fixed-frame or fixed-meter assumptions.

The main limitations are equally consistent. Osu2MIR notes editor limitations, occasional tempo halving or doubling differences across annotators, and the need to exclude densely spaced timing-point regions. The DBN-free beat tracker reports lower continuity metrics and failures on difficult or underrepresented genres. The beat-to-melody system is limited to **monophony** and inherits domain-shift risk from MAESTRO-based training. The chart-generation system is sensitive to BPM and beat-tracking errors and depends strongly on large-scale pretraining. None of these limitations invalidates the corresponding artifact, but they make clear that osu2beat2025 is not a single solved problem; it is a label attached to several research programs centered on rhythmic representation, each with distinct assumptions, outputs, and evaluation criteria.

Source: https://www.emergentmind.com/topics/osu2beat2025