HookTheory Lead Sheet Dataset (HLSD)

• HLSD is a public symbolic corpus of popular music lead sheets containing melodies, chord progressions, key/mode, and sectional labels.
• It offers a rich chord vocabulary with up to 633 distinct chord types and dual formats (JSON and MIDI) for versatile harmonization studies.
• HLSD supports conditional VAE training for melody harmonization and integration with audio for partial structural analysis to enhance boundary detection.

to=arxiv_search.search ाठमाडौं json {"query":"HookTheory Lead Sheet Dataset HLSD arXiv", "max_results": 10} to=arxiv_search.search ＿老司机 一级a做爰片 天天中彩票有output truncated due to platform? to=arxiv_search.search ￣奇米影视 qq天天中彩票 天天彩票怎么 to=arxiv_search.search 凤凰大参考 东臣 ചികിതાઓ ็ตทรู json {"query":"(Wang et al., 2022)", "max_results": 5} to=arxiv_search.search 大发展有限公司官网 ুমեցություն to=arxiv_search.search արզ Rhumela json {"query":"(Chen et al., 2021)", "max_results": 5} The HookTheory Lead Sheet Dataset (HLSD) is a large, public, symbolic lead-sheet collection derived from HookTheory TheoryTabs, where users annotate lead sheets for popular songs. In the research literature, it is characterized as a corpus of melody and chord progressions with additional symbolic annotations such as chord symbols, Roman numerals, key, mode, and section information; it has been used as a training and evaluation resource for melody harmonization and, when linked to audio via time stamps, as a source of partially labeled section excerpts for music structure analysis (Chen et al., 2021, Wang et al., 2022).

1. Definition, provenance, and corpus scope

HLSD is introduced in the cited work as a dataset of lead sheets rather than full symbolic scores. A lead sheet typically contains melody, chord progression, key and mode, and section information; the section usually corresponds to a musically meaningful unit such as chorus, verse, or intro rather than an entire song (Wang et al., 2022). In the harmonization literature, the corpus is described as containing “high-quality and human-arranged melodies with chord progressions,” provided in two formats, event-based JSON files and MIDI files, and including multiple chord label types such as chord symbols and Roman numerals for reference (Chen et al., 2021).

Two distinct operational views of HLSD appear in the literature. In SurpriseNet, HLSD is the sole corpus used for training and evaluating conditional VAE harmonization models, with a predefined split of 17,505 training samples and 500 test samples (Chen et al., 2021). In MuSFA, the open-sourced collection at https://github.com/wayne391/lead-sheet-dataset is used as an aligned audio-symbolic resource containing 18,843 music sections, each with starting and ending times of the section in the corresponding recording (Wang et al., 2022). In that setting, each entry is a section-level excerpt aligned to audio rather than a fully annotated structural description of an entire track.

The dataset’s stylistic orientation is consistently tied to popular music. The MuSFA study describes HLSD as strongly biased toward Western popular music, while SurpriseNet implicitly aligns it with pop-style harmony and common progressions through its modeling choices and evaluation criteria (Wang et al., 2022, Chen et al., 2021). This suggests that HLSD is best understood as a pop-oriented symbolic corpus whose research value lies in the combination of curated harmonic annotations and musically meaningful sectional metadata.

2. Symbolic representation and annotation granularity

A central property of HLSD is the richness of its chord vocabulary. SurpriseNet reports that HLSD contains a total of 633 chord types, including 9th, 11th, 13th, half diminished, and slash chords; earlier HLSD-based studies had simplified the vocabulary to 48 chord types, corresponding to major and minor triads, or 96 chord types, corresponding to major, minor, augment, diminish, suspend, major7, minor7, and dominant7 (Chen et al., 2021). The paper’s contribution is framed explicitly as extending the chord space rather than collapsing complex harmonies into a reduced symbolic inventory.

For model training in SurpriseNet, melody and chords are aligned every two beats in a measure, and chords are encoded into a one-hot format. The experiments use two vocabulary settings, 96 and 633 chord types, with no functional reduction applied in the full-vocabulary case; rare and complex chords remain distinct labels (Chen et al., 2021). The same work also notes a representational limitation: although HLSD contains various rhythmic types such as syncopation and tuplets, the modeling pipeline simplifies chord rhythm and restricts it to two beats in a measure.

The symbolic annotations extend beyond raw chord labels. SurpriseNet notes that HLSD contains rich intonation data, including Roman, symbol, secondary, and mode data, although those annotations are not exploited directly in the reported architecture (Chen et al., 2021). In MuSFA, the important symbolic field is section naming: after parsing metadata from the open collection, the authors derive a taxonomy that includes about 22 different labels, such as chorus, verse, intro, bridge, variation, loop, solo, and combined forms such as verse-and-chorus or intro-and-verse (Wang et al., 2022).

These properties place HLSD between conventional symbolic harmony datasets and structural annotation datasets. It is symbolically dense, especially in chordal detail, but its structural content is frequently local or sectional rather than exhaustive at song level.

3. HLSD in melody harmonization and surprise-conditioned generation

In SurpriseNet, HLSD supports two closely related tasks: a basic melody harmonization task conditioned only on melody, and a surprise-controlled harmonization task conditioned on melody plus a temporal contour derived from chord-transition surprise (Chen et al., 2021). In the basic task, HLSD provides paired melody and chord sequences for learning the conditional distribution $p(\text{chord sequence} \mid \text{melody})$ with a CVAE. The objective is given as

$$\tilde{\mathcal{L}}_{CVAE} = \mathbb{E}[\log p_{\theta}(\mathbf{x}|\mathbf{z},\mathbf{c})] - KL(p_\phi(\mathbf{z}|\mathbf{x},\mathbf{c})||p(\mathbf{z}|\mathbf{c})).$$

Here, $\mathbf{x}$ is the HLSD chord sequence, $\mathbf{c}$ is the conditioning signal, and $\mathbf{z}$ is the latent variable.

The surprise-controlled formulation derives a scalar “surprisingness” value for each chord transition using a first-order Markov chain over the HLSD chord vocabulary:

$$\text{Surprisingness} = -\log p(\text{chord}_t|\text{chord}_{t-1}).$$

This sequence becomes the surprise contour. The raw scalar contour is processed by a Pre-net implemented by a BiLSTM to extend the feature from a scalar to a 256-dimensional vector, which is then concatenated with melody features and latent variables in the conditional generation pipeline (Chen et al., 2021). During training, the contour is computed from the ground-truth HLSD chord sequences; during inference, user-defined contours such as sigmoid, flat, and normal shapes are supplied.

Evaluation on HLSD uses six objective metrics: chord histogram entropy (CHE), chord coverage (CC), chord tonal distance (CTD), chord tone to non-chord tone ratio (CTnCTR), pitch consonance score (PCS), and melody-chord tonal distance (MCTD) (Chen et al., 2021). The reported results show that the vanilla CVAE with $|S|=96$ achieves performance comparable to the state-of-the-art melody harmonization model on harmonicity metrics and obtains the best CTD among the listed systems, while expanding to $|S|=633$ increases chord diversity, as reflected in CHE and CC, but worsens some harmonicity metrics. For surprise control, Spearman’s correlation between the given contour and the realized contour from generated chords is reported as $\rho = 0.517$ with $p < 0.001$ for SurpriseNet, and $$\tilde{\mathcal{L}}_{CVAE} = \mathbb{E}[\log p_{\theta}(\mathbf{x}|\mathbf{z},\mathbf{c})] - KL(p_\phi(\mathbf{z}|\mathbf{x},\mathbf{c})||p(\mathbf{z}|\mathbf{c})).$$0 with $$\tilde{\mathcal{L}}_{CVAE} = \mathbb{E}[\log p_{\theta}(\mathbf{x}|\mathbf{z},\mathbf{c})] - KL(p_\phi(\mathbf{z}|\mathbf{x},\mathbf{c})||p(\mathbf{z}|\mathbf{c})).$$1 for Weighted SurpriseNet (Chen et al., 2021). A plausible implication is that HLSD’s dense chord vocabulary supports controllability by allowing the model to vary harmonic complexity while remaining within the corpus style.

4. HLSD as partially labeled data for music structure analysis

MuSFA repurposes HLSD for a different task family: music structure analysis from audio. In that setting, the key observation is that HLSD does not provide full-track structural annotation in the open collection used by the paper; instead, it provides one labeled section per song together with start and end times in the corresponding recording (Wang et al., 2022). Each padded excerpt is therefore treated as an independent “song” for training, but only one contiguous span inside the excerpt is structurally labeled. The regions before and after that span are unlabeled.

The preprocessing procedure is explicit. For each HLSD entry, the authors sample a random front padding of 8–12 seconds before the section and a random rear padding of 8–12 seconds after the section, enforce a minimum total excerpt length of 30 seconds, and cut the span at the song boundaries if padding would exceed them (Wang et al., 2022). The resulting object is a section-centered audio excerpt in which the interior section has a known structural function label and exact start/end boundaries, while neighboring structure remains unspecified.

To make HLSD compatible with MuSFA’s label space, the 22 HookTheory section labels are mapped into MuSFA structural function classes. The paper gives the following mapping (Wang et al., 2022):

HLSD labels	MuSFA
chorus, chorus-lead-out, theme, verse-and-chorus, theme-recap, pre-chorus-and-chorus	chorus
verse, development, verse-and-pre-chorus, pre-chorus	verse
instrumental, lead-in-alt, lead-in, loop, solo	inst
bridge, variation	bridge
intro, intro-and-chorus, intro-and-verse	intro
outro, pre-outro	outro

The codomain is the common MuSFA label space used across SALAMI-pop, RWC-Pop, Isophonics, and Harmonix; HLSD uses six of those classes, while one more MuSFA label is present in the original system but unused by HLSD (Wang et al., 2022). This mapping is lossy by design. Combined labels such as verse-and-chorus or intro-and-chorus are compressed into a single top-level function, which improves cross-dataset compatibility but removes finer distinctions in sectional interpretation.

5. Training regime, loss masking, and cross-dataset integration

MuSFA predicts two time-varying activation curves from audio: boundary activation and function activation. For fully labeled structural datasets, both losses are computed over all frames. For HLSD excerpts, the boundary targets are defined by the known start and end of the labeled section, but the function loss must be masked outside the labeled span because the padding regions are structurally unknown (Wang et al., 2022). The conceptual change is described as restricting the function loss from

$$\tilde{\mathcal{L}}_{CVAE} = \mathbb{E}[\log p_{\theta}(\mathbf{x}|\mathbf{z},\mathbf{c})] - KL(p_\phi(\mathbf{z}|\mathbf{x},\mathbf{c})||p(\mathbf{z}|\mathbf{c})).$$2

to

$$\tilde{\mathcal{L}}_{CVAE} = \mathbb{E}[\log p_{\theta}(\mathbf{x}|\mathbf{z},\mathbf{c})] - KL(p_\phi(\mathbf{z}|\mathbf{x},\mathbf{c})||p(\mathbf{z}|\mathbf{c})).$$3

so that front and rear padding frames do not contribute to the function-label penalty.

The audio frontend is SpecTNT, and the training configuration is supervised multi-dataset training rather than semi-supervised or unsupervised learning (Wang et al., 2022). HLSD contributes partially labeled excerpts, whereas Harmonix, SALAMI-pop, RWC-Pop, and Isophonics contribute full-track structural annotations. MuSFA’s models operate on fixed-length audio chunks, reported in the paper as MuSFA-24s and MuSFA-36s. The common structural label space allows the function activation network to predict the same top-level categories across all datasets.

This integration strategy addresses a mismatch in data regimes. HLSD contributes many more examples of clear, musically meaningful sections—over 18K excerpts of music sections, compared with 1,563 full songs across the four traditional MSA datasets combined—while the full-song datasets supply global structural context and complete label transitions (Wang et al., 2022). A plausible implication is that HLSD is especially effective for local function recognition and boundary localization, whereas its utility for learning full-form syntax is inherently limited by the absence of complete song-level annotation.

6. Empirical effects, limitations, and recurrent misconceptions

The empirical effect of adding HLSD to MuSFA training is reported consistently across datasets. Using mir_eval, the paper evaluates HR.5F for boundary hit rate with a 0.5-second tolerance, ACC for frame-wise function-label accuracy, CHR.5F for chorus-boundary hit rate at 0.5 seconds, and CF1 for chorus versus non-chorus frame-pair classification (Wang et al., 2022). For MuSFA-24s, the reported changes include Harmonix HR.5F from .570 to .595 and ACC from .701 to .714, SALAMI-pop HR.5F from .490 to .532 and ACC from .544 to .551, RWC-Pop HR.5F from .623 to .643 and ACC from .675 to .677, and Isophonics HR.5F from .590 to .598 and ACC from .550 to .559. The authors summarize the trend as an average improvement of 3\% in boundary detection performance and 1\% in section labeling performance when HLSD is included (Wang et al., 2022).

Two misconceptions recur in discussion of HLSD. The first is that it is a full-song structural annotation dataset. In the MuSFA use case, that is incorrect: the open collection supplies one section per song with start/end times in audio, so the rest of the track is structurally unlabeled (Wang et al., 2022). The second is that HLSD is inherently a simplified triadic harmony dataset. SurpriseNet directly contradicts that reading by emphasizing the 633-chord vocabulary and by treating extended, half-diminished, and slash chords as distinct symbolic categories rather than reducing them to 48 or 96 simplified types (Chen et al., 2021).

The limitations are correspondingly clear. HLSD’s structural annotations are partial rather than exhaustive in the MuSFA setting; single-section labeling prevents direct learning of full sequence patterns or hierarchical form (Wang et al., 2022). The corpus is biased toward Western pop and rock repertoire, which constrains stylistic generalization (Wang et al., 2022). In SurpriseNet, the rich chord vocabulary creates a trade-off: rare chords such as 7th, 9th, 11th, and 13th can degrade melody/chord harmonicity metrics such as CTnCTR, PCS, and MCTD, even while increasing chord diversity (Chen et al., 2021). The paper also notes that rhythmic content in HLSD is richer than the two-beat quantization used in the model, and that Roman numeral, secondary, and mode annotations remain underused (Chen et al., 2021).

Taken together, the literature presents HLSD as a research corpus with two complementary identities. It is, first, a large-scale symbolic lead-sheet dataset with unusually rich chordal detail, suitable for conditional generation, harmonization, and controllable sequence modeling. It is, second, a source of section-level, partially labeled audio excerpts that can materially improve supervised music structure analysis when integrated with full-song annotation datasets. The combination of melody, chord progression, key and mode, and section information makes HLSD distinctive, but the same combination also requires careful attention to annotation scope, label compression, rhythmic simplification, and genre bias (Chen et al., 2021, Wang et al., 2022).