Source-linked AI summary
Gugak-VocalSet: A Multi-Genre Studio-Quality Dataset of Korean Traditional Singing with Sigimsae Annotations
Danbinaerin Han, Wonil Kim, Seah Hong, Dongjoo Moon, Jaeok Lee, Jongpil Lee, Chaewon Kim, Juhan Nam
TL;DR
Notation cannot fully capture sigimsae’s range and variability, and existing Korean vocal resources lack publicly available multi-genre audio with extensive sigimsae labels. Gugak-VocalSet provides such a corpus with audio-grounded annotations and baseline evaluations. Domain-informed F0 representations surpass frozen MERT probing on classification, while pitch and spectral cues play complementary roles in event detection.
Problem
Notation cannot fully capture sigimsae’s range and variability, and existing Korean traditional vocal resources lack publicly available multi-genre audio with extensive sigimsae labels.
Source: §1 · p. 1, §2.1 · p. 2, §4.1 · p. 3Method
Gugak-VocalSet combines controlled recordings of Korean traditional vocal music with a unified sigimsae ontology, audio-grounded annotations, and classification and event-detection baselines.
Source: §1 · pp. 1–2, §3.2 · p. 2Results
Domain-informed F0 input surpasses frozen MERT probing on classification, while F0 captures ornament identity, mel captures event boundaries, and EDTCN remains competitive with MERT on event detection.
Source: §1 · p. 2Takeaways & Limitations
Controlled recordings with cross-genre annotations enable systematic corpus-level comparative analysis of Korean traditional vocal music.
Source: §7 · p. 6The ontology prioritizes cross-genre consistency over within-genre granularity, absorbing some genre-specific techniques and mode distinctions into pitch-contour-based parent categories.
Source: §6 · p. 6
Abstract
from arXiv · showhide
We present Gugak-VocalSet, the first publicly available audio dataset for Korean traditional singing with dense annotations of sigimsae, a structured pitch contour system of vocal expressions. Gugak-VocalSet contains 24.5 hours of unaccompanied solo vocal recordings across 305 tracks of representative repertoire curated by the National Gugak Center, performed by 20 professional singers in a controlled studio environment with fixed key and tempo. In addition to genre, lyrics, and mood/timbre tags, the dataset provides a unified sigimsae ontology and approximately 68,000 segments (71,700 instances including multi-label) of audio-grounded sigimsae annotations; the ontology adopts a functional-position-based hierarchy designed for consistent application across genres, and 14 expert annotators specializing in Korean traditional music performed segment-level labeling accordingly. Baseline experiments on sigimsae classification and event detection show that domain-informed F0 representations remain competitive with, and on classification surpass, frozen probing of a general-purpose music foundation model, while pitch and spectral cues play complementary roles in event-level detection. Companion page: https://www.danbinaerin.com/Gugak-VocalSet/
1. INTRODUCTION
Sigimsae’s expressive pitch cannot be fully represented in notation, and rigorous performance analysis therefore requires audio-grounded study. Gugak-VocalSet addresses this gap with a cross-genre vocal corpus, annotations, and baseline experiments.
1. INTRODUCTION
- Sigimsae shapes melodic identity and, alongside jangdan and garak, is regarded as one of Korean music’s three fundamental elements.Pieces using the same scale may belong to different tori when their sigimsae differ.Source: §1 · p. 1
- Notation reduces continuous sigimsae to symbolic approximations shaped by transcriber interpretation and cannot capture the full range and variability of its realization.The paper argues that rigorous analysis of actual performance therefore requires working directly from audio.Source: §1 · p. 1
- Gugak-VocalSet is a controlled-studio corpus of 305 tracks (24.5 hours) of unaccompanied solo vocals across Korean traditional music genres, with approximately 68,000 audio-based sigimsae annotations.Source: §1 · pp. 1–2
- Domain-informed F0 input surpasses frozen MERT probing on sigimsae classification, while pitch and mel cues capture complementary aspects of event detection and EDTCN remains competitive.The reported event-detection roles are ornament identity for F0 and event boundaries for mel.Source: §1 · p. 2
2. RELATED WORK
Existing Korean traditional music resources leave gaps in public access, genre coverage, and sigimsae labeling, while prior expressive-pitch datasets address other traditions or narrower vocabularies. Gugak-VocalSet contributes multi-genre Korean vocal recordings and event-level sigimsae annotations.
2.1 Non-Western Music Datasets
- No existing dataset targets Korean traditional vocal music with openly available multi-genre recordings: one corpus is Korea-only with six sigimsae categories, and another covers one genre without distributing audio.Gugak-VocalSet fills this gap with openly available recordings.Source: §2.1 · p. 2
2.2 Vocal Technique and Expressive Pitch Analysis in MIR
- Gugak-VocalSet extends event-level expressive-pitch analysis to Korean traditional music with a cross-genre sigimsae ontology and large-scale annotations.Earlier work includes clip-level singing-technique catalogues, timestamped J-POP technique detection, and expert annotations for six Hindustani gamaka types.Source: §2.2 · p. 2
- Figure 1 presents the dataset’s genre hierarchy and the number of tracks in each category.It provides a view of how track counts are distributed across categories; the passage gives no specific category counts.Source: Figure 1 · p. 2
3. GUGAK-VOCALSET
Gugak-VocalSet covers diverse Korean traditional vocal repertoires through curated tracks, controlled studio recordings, and layered metadata and annotations. The recording choices support corpus-level comparison, while the dataset preserves original and refined audio.
3.1 Korean Traditional Vocal Music
- Korean traditional vocal music varies across courtly and popular traditions and regional or school-based transmission, including differences in style, vocal production, and sigimsae.Pansori and Namdo folk songs are described as sharing southern-region scalar and vocal characteristics.Source: §3.1 · p. 2
3.2 Data Configuration
- The dataset contains 305 tracks of representative repertoire curated by the National Gugak Center, with subgenre labels encoding narrative or regional origin.For pansori, tracks reflect singers trained in the byeongchang tradition rather than full-length pansori.Source: §3.2 · p. 2
3.3 Audio Recording & Refinement
- Recordings are unaccompanied solo vocals by 20 professional singers, made at fixed tonic and tempo for corpus-level comparison and distributed as original and refined audio.The refined take was selected from two takes and underwent mixing, pitch tuning, tempo correction, and two-stage quality review.Source: §3.3 · p. 3
3.4 Annotations & Metadata
- Track metadata includes genre, singer, jangdan, tempo, key, bilingual context captions, and expert-curated mood and timbre tags.Lyrics are aligned at the phrase level according to jangdan, and the annotations include sigimsae regions; Figure 2 groups 17 sigimsae types by functional position and shows their symbols, contours, and descriptions.Source: §3.4 · p. 3, Figure 2 · p. 3
4. SIGIMSAE ANNOTATION
The section defines a cross-genre sigimsae ontology organized by functional position in the melodic line, then describes its expert annotation and review. The released annotations cover 71,742 instances, while a focused consistency check finds stronger agreement on event locations than on fine-grained types.
4.1 Unified Sigimsae Ontology
- The ontology addresses cross-genre inconsistencies by providing a unified system applicable regardless of genre or transmission method.Its top-level hierarchy uses functional position in the melodic line.Source: §4.1 · pp. 3–4
- The ontology organizes 17 sigimsae types by functional position—onset, mid-note, offset, and transition—while excluding genre-specific production and timbre distinctions.Its types were refined through score surveys and a prototype annotation round; traditional names were retained when unambiguous, with intuitive Korean names and contour-reflecting symbols for others.Source: §4.1 · p. 4
4.2 Annotation Process and Statistics
- Fourteen specialists annotated sigimsae and phrase-level lyric–audio alignment, followed by a two-stage review with reannotation guidance and reviewer corrections.Pieces were assigned by annotators’ genre specialization, and guidelines included audio examples and pitch-contour visualizations.Source: §4.2 · p. 4
- 71,742 instances across 68,272 segments in 305 tracks include 3,453 multi-label segments (5.1%); overall mean segment duration was 0.47 s.Figure 4 plots counts per type and duration distributions; vibrato types have a median duration of 0.76 s, longer than other ornaments.Source: §4.2 · p. 4, Figure 4 · p. 4
- Review yielded EvF1@200ms = .903 and onset F1 = .910; 93.9% of time-agreed events retained their 17-class label, while region count changed by 1.5%.The authors describe review as targeted correction, with released labels largely reflecting the 14 specialist annotators’ work.Source: §4.2 · p. 4
- Class-agnostic EvF1@200ms was .597, versus class-aware EvF1@200ms of .357; merging confused types into 10 classes raised same-label agreement from 56% to 78%.The focused check covered the first 60 s of 10 tracks, independently annotated by two project members.Source: §4.2 · p. 5
5. BASELINE EXPERIMENTS
The baselines compare domain-informed pitch and spectral inputs with frozen music-model representations across sigimsae classification and event detection. F0CNN leads the reported classification comparison, while pitch and mel inputs show complementary strengths in detection.
5.1 Experimental Setup
- Evaluation uses song-level five-fold splits; 5.1% multi-label segments are reduced for yoseong-dominant cases or skipped for other ambiguous combinations.Classification compares F0 and mel task-specific CNNs with frozen foundation-model probes; full multi-label handling is left for future work.Source: §5.1 · p. 5, §5.2 · p. 5
5.2 Task 1: Sigimsae Classification
- F0CNN reaches Macro F1 .441, surpassing MelCNN (.400) and frozen MERT-v195M (.358) on 17-category classification.The MERT-v195M score is at its best probing layer, L7.Source: §5.2 · p. 5, Table 2 · p. 5
- All three foundation models perform worst at L12 and peak at intermediate layers L5–L7, consistent with pitch-level ornament information being more distributed there.The paper relates this pattern to prior MIR observations about higher-level semantic information concentrating in upper MERT-family layers [36].Source: §5.2 · p. 5
- 78.4% vs. 78.0% same-label rates (239 vs. 238 of 305) show that model-confusion and IAA-derived groupings recover similar coarse groupings.Prominent F0CNN confusions include yoseong subtypes and Onset-DFl versus Onset-UTn; the paper interprets fine-level disagreement as acoustic ambiguity rather than annotation noise.Source: §5.2 · p. 5, Figure 5 · p. 6
- Cross-genre Macro F1 drops 34.0% for J→M and 19.2% for M→J, with cross Acc ≈.43; gender is fully confounded with genre.The comparison matched jeongak and minsogak at 102 pieces; the authors note that the asymmetry may reflect genre variability or register differences.Source: §5.2 · pp. 5–6, Table 3 · p. 5
5.3 Task 2: Sigimsae Event Detection
- Ev.F1@200 .763 vs. .756 and IoU .712 vs. .706 show EDTCN(mel) matching or slightly exceeding MERT-v195M on event metrics at <2% of its parameter cost.F0 leads on frame-level Macro F1 (.324 vs. .269 for mel and .311 for MERT-v1), while mel leads on event-level metrics.Source: §5.3 · p. 6, Table 4 · p. 6
- F0 captures ornament identity, while mel cues such as energy envelope and timbral transitions likely help localize events.This distinction is notable because sigimsae is defined by pitch contour.Source: §5.3 · p. 6
6. DESIGN TRADE-OFFS
The design trade-offs prioritize reproducible cross-genre analysis while constraining expressive and within-genre detail.
6. DESIGN TRADE-OFFS
- Fixed tonic and tempo constrain temporal and tonal flexibility integral to traditional Korean vocal expression; some singers reported the constraint felt unnatural.The authors treat this as a necessary trade-off for reproducible corpus-level analysis.Source: §6 · p. 6
- The recording constraint may limit captured expressive freedom, while the ontology’s broader categories trade detail for systematic cross-genre comparison.The authors describe the recording protocol as a foundation for future work with more naturalistic recordings.Source: §6 · p. 6
- The ontology prioritizes cross-genre consistency over within-genre granularity, absorbing genre-specific vocal-production and mode distinctions into pitch-contour parent categories.Annotators less familiar with a genre may default to the closest general type; finer-grained genre-internal extensions remain open for future work.Source: §6 · p. 6
7. APPLICATIONS & CONCLUSIONS
Gugak-VocalSet combines controlled cross-genre recordings with a functional-position-based sigimsae ontology and audio-grounded annotations. The dataset supports corpus-level comparative analysis and a range of analytical, educational, and generative applications.
7. APPLICATIONS & CONCLUSIONS
- Domain-informed pitch representations remain competitive against general-purpose music foundation models, while the ontology demonstrates discriminative capacity.The corpus contains 24.5 hours of Korean traditional vocal music and approximately 68,000 audio-grounded annotations.Source: §7 · p. 6
- Controlled-environment recordings with cross-genre annotations enable systematic corpus-level comparative analysis of Korean traditional music.Gugak-VocalSet provides a unified, functional-position-based sigimsae ontology.Source: §7 · p. 6
- The annotations support gugak archive analysis, singer or genre identification, performance-education feedback, and controlled singing synthesis that preserves sigimsae fidelity.The paper frames these as potential analytical, interactive, and generative applications.Source: §7 · p. 6