HaitongSUN/prosody-abx

Prosodic ABX Speech Dataset

This dataset was collected for Prosodic ABX: A Language-Agnostic Method for Measuring Prosodic Contrast in Speech Representations (arXiv:2604.02102). The paper introduces prosodic ABX, an ABX-like framework to evaluate prosodic contrast encoding in speech representations.

The release contains clean read-speech recordings for English and Japanese prosodic minimal pairs:

• English stress: noun/verb lexical stress contrasts in English target words.
• Japanese pitch accent: pitch-accent contrasts in Japanese target words or short expressions.

To our knowledge, this is the first public dataset focused on clean sentence recordings of lexical-prosody minimal pairs for both English and Japanese. For the full task definition, evaluation setup, and experimental details, see the paper: https://arxiv.org/abs/2604.02102.

Files

The uploaded release is intended to contain only the two parquet files.

File	Description
english_stress.parquet	English stress subset with full sentence audio and target timestamps.
japanese_pitch_accent.parquet	Japanese pitch-accent subset with full sentence audio and target timestamps.

Collection and Annotation

Speakers read carrier sentences designed to contain lexical-prosody contrasts. The English corpus contains stress minimal pairs; the Japanese corpus contains pitch-accent minimal pairs. For each pitch-accent minimal pair, we constructed two pairs of carrier sentences with identical context for both words.

Target-word boundaries were first obtained with Montreal Forced Aligner (MFA) word-level alignment. The MFA boundaries were then manually checked and adjusted where necessary. The released target_onset and target_offset fields are the final boundaries used for the dataset.

English stress labels were manually annotated by a native American English speaker. Japanese pitch accent labels were checked by a native Japanese speaker.

Column Descriptions

English Stress

Column	Type	Description
id	string	Example ID, prefixed with EN_.
audio	audio	Full 48 kHz sentence audio.
speaker	string	Language-prefixed speaker ID, e.g. EN_S003.
text	string	Reading sentence.
target	string	English target word.
target_onset	float	Target onset in seconds, relative to the full sentence audio.
target_offset	float	Target offset in seconds, relative to the full sentence audio.
label	integer	English stress label used for the prosodic contrast.
context_set	string	English carrier-sentence set, A or B.
lexical_category	string	Lexical category of the target use, noun or verb.

Japanese Pitch Accent

Column	Type	Description
id	string	Example ID, prefixed with JP_.
audio	audio	Full 48 kHz sentence audio.
speaker	string	Language-prefixed speaker ID, e.g. JP_S009.
text	string	Reading sentence.
target	string	Japanese target surface form.
surface_kana	string	Kana sequence for the target.
target_onset	float	Target onset in seconds, relative to the full sentence audio.
target_offset	float	Target offset in seconds, relative to the full sentence audio.
label	integer	Manually assigned Japanese pitch-accent label. 0 indicates no pitch fall; positive values indicate the mora after which the pitch falls, e.g. 1 means a fall after the first mora.
context_id	integer	Shared context-pair ID for adjacent Japanese prosodic minimal-pair sentences.

Citation

If you use this dataset, please cite:

@article{sun2026prosodic,
  title={Prosodic ABX: A Language-Agnostic Method for Measuring Prosodic Contrast in Speech Representations},
  author={Sun, Haitong and McIntosh, Stephen and Choi, Kwanghee and Yeo, Eunjung and Saito, Daisuke and Minematsu, Nobuaki},
  journal={arXiv preprint arXiv:2604.02102},
  year={2026}
}

License

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