mattwinwood/vocal-score-synthetic-smoke-test
Vocal Score Synthetic Smoke Test A tiny, fully synthetic fixture for smoke-testing monophonic vocal-to-notation pipelines. It contains one ten-second “ah”-like synthesized rendition of the public-domain melody commonly known as “Twinkle, Twinkle, Little Star,” its scripted note sequence, and observed exports from my Vocal Score research pipeline. No human voice, copyrighted recording, voice embedding, lyric, or personal data is included. Files twinkle.wav — 22.05… See the full description on the dataset page: https://huggingface.co/datasets/mattwinwood/vocal-score-synthetic-smoke-test.
Vocal Score Synthetic Smoke Test
A tiny, fully synthetic fixture for smoke-testing monophonic vocal-to-notation pipelines. It contains one ten-second “ah”-like synthesized rendition of the public-domain melody commonly known as “Twinkle, Twinkle, Little Star,” its scripted note sequence, and observed exports from my Vocal Score research pipeline.
No human voice, copyrighted recording, voice embedding, lyric, or personal data is included.
Files
twinkle.wav— 22.05 kHz mono synthetic vowel-like tones;metadata.jsonl— one row with generation parameters and scripted reference notes;twinkle.mid— observed Vocal Score MIDI export;twinkle.musicxml— observed editable notation export;twinkle.pdf— observed rendered score;twinkle-pianoroll.png— observed piano-roll visualization; andmake_fixture.py— deterministic source generator for the input audio.
Reference versus observed output
The reference_midi_pitches and reference_note_beats fields come from the generator and are ground truth for the synthesized input. The MIDI, MusicXML, PDF, and piano-roll files are observed pipeline outputs, not ground-truth labels and not evidence of transcription accuracy.
Example loading
from datasets import load_dataset
dataset = load_dataset("mattwinwood/vocal-score-synthetic-smoke-test", split="train")Intended use
- installation and end-to-end smoke tests;
- verifying that an audio-to-notation pipeline emits expected artifact types;
- regression debugging for tempo, quantization, MIDI, MusicXML, and rendering stages; and
- teaching examples where redistributable audio matters.
Limitations
- One melody is not a benchmark.
- The signal is far cleaner and simpler than a human vocal.
- The fixture does not test lyrics, breath noise, accompaniment, timing drift, vibrato diversity, or polyphony.
- Do not use the observed exports as proof of production accuracy.
License
CC BY 4.0. Attribution: Matt Winwood, “Vocal Score Synthetic Smoke Test.” The underlying traditional melody is public domain.
