---
pretty_name: WaveMaster External Vibrato Evaluation Set
language:
- it
license: other
task_categories:
- audio-classification
tags:
- audio
- singing-voice
- vibrato
- music-information-retrieval
size_categories:
- n<1K
---

# WaveMaster External Vibrato Evaluation Set

This dataset supports the external-corpus analysis in *Presence Is Not
Placement: Auditing What Vibrato Metrics Measure*. It contains unaccompanied
recordings of eight professional singers performing *Caro mio ben*. A full
release may additionally include the counterfactual renders used to test
whether a learned vibrato-presence judge responds to note-level vibrato
placement; `metadata.csv` is the authoritative list of included files.

## Consent and identifiability

The singers participated through the Bach-Millennium Program. Each participant
gave written informed consent before recording, including consent to
redistribute their voice recording. A singing voice is inherently identifiable;
the singer IDs in this repository are pseudonyms, not a claim that the audio is
anonymous. Names, contact information, consent forms, and unrelated recordings
are not included.

The recordings were used only as an external evaluation set. They were not used
to train WaveMaster or any of its five fold-specific classifiers.

## Contents

Every release contains the eight original recordings:

```text
audio/
  original/singer_01.mp3 ... singer_08.mp3
metadata.csv
```

If redistribution of modified derivatives is also authorized, the builder can
add the exact audit renders:

```text
audio/
  identity/singer_01.wav ... singer_08.wav
  shuffle/singer_01.wav ... singer_08.wav
  jitter/singer_01.wav ... singer_08.wav
  depth_0p5/singer_01.wav ... singer_08.wav
  depth_2/singer_01.wav ... singer_08.wav
  constant_130/singer_01.wav ... singer_08.wav
  constant_420/singer_01.wav ... singer_08.wav
```

`original` contains the contributed performances. When included, `identity` is a WORLD
round-trip with no intended pitch change. `shuffle` reassigns the measured
vibrato depths among notes. `jitter` applies a similarly sized random note-level
perturbation. `depth_0p5` and `depth_2` globally halve or double vibrato depth.
`constant_130` and `constant_420` assign the same target depth, in cents, to
every measured note.

`metadata.csv` records the pseudonymous singer ID, condition, manipulation
amount, source-versus-derived status, duration, sample rate, channel count,
codec, file size, and SHA-256 checksum.

Only the *Caro mio ben* files used in the paper are included. The singers also
recorded scales, passaggio exercises, and other songs; those recordings are not
part of this research release.

## Dataset creation

The original recordings were made in a room and with equipment different from
VocalSet. When included, the counterfactual files were generated with the
paper's public analysis code. Fundamental frequency was decomposed into a slow trend, a
3--9 Hz vibrato component, and a residual. Only the vibrato component was
scaled or reassigned, after which WORLD resynthesized the audio while retaining
the analyzed spectral envelope and aperiodicity parameters.

## Intended use and limitations

The intended use is reproduction and extension of research on vibrato
measurement, expressive-placement evaluation, and singing synthesis. The set is
small: eight singers, one Western classical song, one take per singer, and one
recording context. It should not be treated as representative of singers,
singing traditions, or listeners generally. Counterfactual files contain WORLD
resynthesis artifacts and are controlled stimuli, not natural performances.

Do not infer identity or sensitive traits from the recordings. Do not use the
recordings to impersonate a participant. See [TERMS.md](TERMS.md) for the
applicable release terms.

## Citation

Citation information will be added when the paper has a stable public URL.

## Contact

Questions about the dataset can be opened as a discussion on its Hugging Face
repository or sent to the corresponding author listed in the paper.