VocalGrad

VocalGrad is an audio benchmark for evaluating whether a model can detect the direction of gradual perceptual change in speech. This public release contains the test split only.

Each example contains one audio clip and one target attribute. The task is to answer whether that attribute increases or decreases over time.

Code: ynklab/vocalgrad.

Task

Given an audio clip and an attribute name, predict one of two labels:

• increase
• decrease

The ground-truth label is derived from the metadata field direction:

• direction=up -> increase
• direction=down -> decrease

Public Release Contents

• split: test
• categories: 9
• source clips per category: 100
• combinations per source clip: 24
• factorization: 3 tiers x 4 curves x 2 directions
• generated clips per category: 2400
• total audio clips: 21600

The deprecated voice_brightness category is not included.

Hugging Face Audio Layout

This repository follows the Hugging Face audio dataset convention:

README.md
metadata.csv
audio/00000/
audio/00001/
...
metadata/summary.json

• metadata.csv is located at the repository root.
• metadata.csv contains a file_name column.
• file_name contains full relative paths to benchmark clips under audio/.
• audio/ contains only benchmark clips referenced by metadata.csv.
• Audio files are sharded into numbered subdirectories to stay below repository directory file-count limits.

Metadata Columns

Common columns include:

• file_name
• category
• benchmark_clip_id
• source_text
• attribute
• prompt
• label
• direction
• curve
• tier
• source_clip_id
• speaker_id
• sentence_id
• duration_sec
• onset_sec
• offset_sec
• split

Categories

• volume
• speaking_speed
• voice_pitch
• background_noise
• audio_distortion
• audio_roughness
• voice_clarity
• voice_vibration
• echo

Prompting

Use the following prompt structure:

Does the {attribute} increase or decrease over time?

Answer with only one word: "increase" or "decrease".

Current category-to-attribute mapping:

• volume -> volume
• speaking_speed -> speaking speed
• voice_pitch -> voice pitch
• background_noise -> background noise
• audio_distortion -> audio distortion
• audio_roughness -> audio roughness
• voice_clarity -> voice clarity
• voice_vibration -> voice vibration
• echo -> echo

Evaluation

Normalize model output to lowercase text and match the following whole words using word boundaries:

• increase or increases -> increase
• decrease or decreases -> decrease

Accept a prediction only when exactly one direction is matched. Outputs matching both directions or neither direction, missing responses, and failed generations are counted as incorrect.

Main metric:

accuracy = (# correct predictions) / (# total samples)

The denominator includes all evaluated clips with valid ground-truth labels, including those with invalid or missing predictions.

Recommended breakdowns:

• overall
• per category
• per tier
• per curve
• per direction

License

VocalGrad is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

The source speech clips are derived from the CSTR VCTK Corpus version 0.92, which is distributed under the Creative Commons Attribution 4.0 International Public License.

Source Dataset Credit

VocalGrad uses source speech from the CSTR VCTK Corpus:

Yamagishi, Junichi; Veaux, Christophe; MacDonald, Kirsten. (2019). CSTR VCTK Corpus: English Multi-speaker Corpus for CSTR Voice Cloning Toolkit (version 0.92), [sound]. University of Edinburgh. The Centre for Speech Technology Research (CSTR). https://doi.org/10.7488/ds/2645.

Original dataset record: https://datashare.ed.ac.uk/handle/10283/3443

License: Creative Commons Attribution 4.0 International https://creativecommons.org/licenses/by/4.0/

Modifications

The audio clips in VocalGrad are transformed derivatives of VCTK source clips. VocalGrad applies controlled time-varying transformations to create benchmark examples for detecting whether a target perceptual attribute increases or decreases over time. The included metadata identifies the source clip ID and the transformation condition for each derived clip.