Dataset Card for "CryCeleb2023"

Dataset Summary

The CryCeleb2023 dataset is a compilation of cries gathered from 786 infants from various hospitals.
The 26k audio files make up 6.5 hours of pure expiration sounds.
The dataset also contains information on the time of recording, which is either within the first hour(s) of life or
upon hospital discharge, typically within 24 hours of birth.

Supported Tasks and Leaderboards

CryCeleb2023 competition

Dataset Structure

Audio folder contains short wav files (16 kHz wav PCM).

audio - folder with audio files structured by infant ID

audio/
  train/
    spk1/
      B/
        spk1_B_001.wav
        ...
        spk6_B_001.wav
      ...
      D/
        spk1_D_001.wav
        ...
    ...
    spk586
      ...
  dev/
    ...(similar to train)...
  test/
    anonymous1/
      B/
        ...

In this folder structure:

• spkN: folder with recordings corresponding to baby N
• B/D: time of recording (birth or discharge)
• 001, 002,, etc - chronological index of cry sound (expiration)

metadata.csv - metadata associated with each audio file

dev_pairs.csv - pairs of birth/discharge recordings used for evaluating development set (available to challenge participants)

test_pairs.csv - pairs of birth/discharge recordings used in CryCeleb2023 evaluation (public and private scores)

Data Instances

Audio files 16 kHz wav PCM - manually segmented cry sounds (expirations)

Data Splits

Number of Infants by Split and Time(s) of Recording(s)

Time(s) of Recording	train	dev	test
Both birth and discharge	348	40	160
Only birth	183	0	0
Only discharge	55	0	0
	586	40	160

Source Data

Audio recordings of infant cries made by android application

Annotations

Annotation process

• Manual segmentation of cry into three categories: expiration, inspiration, no cry
• Only expirations kept in this corpus
• Manual review to remove any PIIs

Personal and Sensitive Information

PII such as intelligible background speech, etc, were removed from the data.

All identities are also anonymized.

Considerations for Using the Data

Discussion of Biases

The dataset only covers infants born in one country

Other Known Limitations

Dataset only includes expirations.

Recording quality varies

Additional Information

Dataset Curators

Ubenwa.ai (contact: [email protected])

Licensing Information

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License

Citation Information

Please cite the following paper if you use this dataset

@article{ubenwa2023cryceleb,
      title={CryCeleb: A Speaker Verification Dataset Based on Infant Cry Sounds}, 
      author={David Budaghyan and Charles C. Onu and Arsenii Gorin and Cem Subakan and Doina Precup},
      year={2023},
      journal={preprint arXiv:2305.00969},
}