Dataset Card — VocalSet-Breath

Dataset summary

VocalSet-Breath is a hand-labeled annotation layer on top of VocalSet (Wilkins et al., 2018) that adds time-aligned audible-breath-event labels. To our knowledge it is the first openly released dedicated breath-event layer for singing voice — onset/offset times with per-event confidence and explicit hard negatives. (Singing corpora such as GTSinger, Opencpop, and M4Singer carry breath only as phoneme tokens or unvoiced-region boundaries; Nakano et al. 2008 hand-labeled 1,488 singing breaths that were never released.) Current release: 250 clips · 215 audible breath events · 407 hard negatives — 118 song excerpts carry nearly all breaths; 132 technique exercises are verified breath-free negative singing.

Languages

English. (VocalSet base corpus is English vowel exercises + English-text excerpts.)

Supported tasks

• Breath-event detection in singing voice (frame-level or event-based)
• Phrase segmentation in singing voice
• Hard-negative learning — explicit fricative / lip-noise / /h/ negatives for precision-aware training
• Cross-corpus evaluation when combined with speech-domain breath corpora (Respiro-en, etc.) or with GTSinger

Source data

• Audio: VocalSet (Wilkins et al., ISMIR 2018). 10.1 hours, 20 singers, 17 vocal techniques. Not redistributed here — fetch from Zenodo (DOI: 10.5281/zenodo.1193957).
• Annotations (this work): hand-labeled by Festus Ewakaa Kahunla following the labeling rubric. Pre-filled from a neural breath detector ("v11") and reviewed clip-by-clip.

Annotations

Per .breath.json file:

Field	Type	Description
audio_file	string	VocalSet wav filename
sample_rate	int	16000 (resampled)
duration_sec	float	clip duration
labeler	string	who labeled
label_date	string	YYYY-MM-DD
tool_version	string	labeling tool version
review_time_sec	int	seconds spent labeling this clip
pre_fill_source	string	model checkpoint that produced draft markers
breath_events	list	confirmed audible inhalations
silent_breaths	list	v1 erratum: these entries predominantly mark silent pauses, not confirmed inaudible breaths. Treat as silence/uncertain regions, never as positives. Semantics will be corrected in v2.
uncertain	list	low-confidence events
hard_negatives	list	explicit negatives (fricative, lip noise, etc.)
exhales	list	audible outward breaths
notes	string	free-form clip-level notes

Each event has start_sec, end_sec, optional confidence (high/medium/low), optional notes.

Splits

By-singer splits (no leakage). Singer IDs in splits/{train,val,test}.txt. Generated reproducibly via tools/make_splits.py with a locked seed.

Annotation process

1. Audio loaded into in-browser labeling tool with pre-filled markers from a neural breath detector.
2. Labeler reviews each marker (accept / delete / nudge) using spectrogram + waveform + audio playback.
3. Labeler adds missed breaths, marks hard negatives, tags uncertain cases.
4. Quality discipline: anti-drift practices (calibration, edge-case log) and inter-annotator agreement on a ≥15% subset (Cohen's κ + event-F1). See labeling rubric §5.

Considerations

• Bias / scope: VocalSet is English vowel exercises + short song excerpts performed by 20 professional singers in studio conditions. It is not representative of amateur singing, popular music with accompaniment, non-English languages, or in-the-wild recording conditions. Models trained on it will need cross-corpus eval (e.g., on GTSinger or new wild recordings) to validate generalization.
• Annotator effect: single primary labeler; the IAA subset partially mitigates but does not eliminate annotator bias. Future versions will add additional labelers.
• License of base data: VocalSet is CC BY 4.0 and Wilkins et al. 2018 must be cited.

Citation

See CITATION.cff.

Maintainer

Festus Ewakaa Kahunla, Drexel University.