Datasets:

Alcadramin
/

signalign-bench

license: other
license_name: per-track-cc
license_details: >-
  Word timing annotations derive from JamendoLyrics and carry each song's
  original Creative Commons license (BY / BY-SA / BY-ND / BY-NC / BY-NC-SA /
  BY-NC-ND — see the `license` field per record). Difficulty bucket annotations
  (buckets.csv) are original work released under CC-BY-4.0. Audio is NOT
  redistributed here; fetch it from the jamendolyrics dataset.
task_categories:
  - automatic-speech-recognition
language:
  - en
  - fr
  - de
  - es
tags:
  - music
  - lyrics
  - alignment
  - singing
  - benchmark
  - evaluation
pretty_name: signalign-bench
size_categories:
  - n<1K

signalign-bench v0.1

The first difficulty-bucketed word-level lyrics alignment benchmark for sung vocals.

Existing lyrics-alignment evaluation reports one blended number. Singing fails aligners in different ways — fast rap, melisma, held notes, vocals buried in the mix — and a single number hides all of it. This benchmark tags every track with its dominant difficulty and reports metrics per bucket.

Maintained at: https://github.com/alcadramin/signalign (evaluation harness, baselines, protocol).

File	What
`jamendolyrics.jsonl`	79 tracks, word-level gold onsets/offsets, difficulty bucket, language, per-track license
`buckets.csv`	The raw human bucket annotations (id → difficulty)
`results/`	Baseline system scores (raw JSON, per bucket)

Schema per record: id, audio_path (relative to the jamendolyrics dataset root), duration, words[{text,start,end}], difficulty, language, source, license.

Difficulty buckets

Assigned by ear by a human annotator (single annotator, v0.1 — annotator: alca, 2026-08-14). One dominant label per track; ties broken by fixed priority rap > melisma > held > buried.

bucket	tracks	meaning
clean	46	clear vocal, plain delivery
buried	13	vocal low in the mix / heavy accompaniment
rap	11	fast dense delivery
melisma	5	multi-pitch syllables, vocal runs
held	4	delivery built from long sustained notes

Audio

Not re-hosted. Word timings and buckets reference the jamendolyrics/jamendolyrics dataset (same track ids / file names):

hf download jamendolyrics/jamendolyrics --repo-type dataset --local-dir data/jamendolyrics

Note: top-level mp3/ entries in that download may materialize as text pointer files; real audio lives under subsets/<lang>/mp3/.

Metrics protocol

Defined and implemented in eval/score.py (single source of truth): per-bucket coverage, MAE, MedAE (matched words, ms), PCO@{100,200,300}ms with all gold words as denominator (a dropped word counts as a miss). Word matching = LCS over normalized text.

Baseline results (v0.1)

Demucs htdemucs vocal stem → forced alignment. Full tables in results/.

system	lyrics input	coverage	MAE	MedAE	PCO@100
wav2vec2-base-960h (speech) + gold lyrics	yes	1.00	150ms	41ms	79.7%
WhisperX medium (ASR path)	no	0.56	2187ms	54ms	38.9%

Per-bucket highlights: rap fails catastrophically not gradually (42ms MedAE, 214ms MAE, lyrics-informed); melisma smears onsets (worst PCO@100); held notes are the easiest bucket, not the hardest; WhisperX drops 44% of sung words. Whisper language ID misdetected a German rap track as Khmer.

Limitations

Single annotator, track-level buckets (v0.2 targets segment-level).
79 tracks, 4 European languages; no extreme-technique or a-cappella-rap hard cases yet (planned additions).
Baseline alignments use an English character vocabulary after ASCII folding for all languages.

Citation

@misc{signalign-bench,
  title = {signalign-bench: a difficulty-bucketed word-level lyrics
           alignment benchmark for sung vocals},
  author = {alca},
  year = {2026},
  url = {https://huggingface.co/datasets/Alcadramin/signalign-bench}
}

Please also cite the underlying JamendoLyrics dataset (Stoller et al.; Durand et al., see the jamendolyrics dataset card).