ArlingtonCL2/Canine-Age-Transition-Vocalization-Dataset

Canine Age Transition Vocalization Dataset

Paper: A Data-driven Approach to the Longitudinal Study of Canine Vocal Pattern Development

Dataset: https://huggingface.co/datasets/ArlingtonCL2/Canine-Age-Transition-Vocalization-Dataset

Dataset Summary

The Canine Age Transition Vocalization Dataset is a longitudinal canine vocalization dataset designed to support research on how dog vocal patterns develop across the lifespan.

The dataset contains 79,142 Bark Units (BUs) extracted from 55,718 Bark Sequences (Barkseqs), totaling 11.4 hours of vocalizations from 125 individual dogs across 6 breeds. Each dog is associated with curated longitudinal metadata, including breed, age in months, and developmental age group.

The released dataset includes Bark Unit audio, parent Barkseq audio, anonymized metadata, temporal alignment between Bark Units and Barkseqs, and Elemental Dog Bark Unit (EDBU) transcripts.

File Structure

The audio is distributed as tar shards to avoid storing more than 100,000 loose files in the repository.

Canine-Age-Transition-Vocalization-Dataset/
├── README.md
├── barkunit_metadata.tsv
└── shards/
    ├── audio_barkunits_dogs_000001_000025.tar
    ├── audio_barkunits_dogs_000026_000050.tar
    ├── audio_barkunits_dogs_000051_000075.tar
    ├── audio_barkunits_dogs_000076_000100.tar
    ├── audio_barkunits_dogs_000101_000125.tar
    └── audio_barkseqs_all.tar

The paths stored in barkunit_metadata.tsv refer to paths inside the extracted shard contents, for example:

audio_barkunits/dog_000001/...
audio_barkseqs/dog_000001/...

Metadata

barkunit_metadata.tsv contains one row per Bark Unit.

Column	Description
barkunit_audio	Relative path to the Bark Unit audio file inside the extracted shards.
barkseq_audio	Relative path to the parent Barkseq audio file inside the extracted shards.
has_barkseq_audio	Whether parent Barkseq audio is available.
dog_id	Anonymized dog identifier.
video_id	Anonymized source video identifier.
barkseq_id	Anonymized Barkseq identifier.
bark_unit_id	Anonymized Bark Unit identifier.
breed	Dog breed.
age_months	Dog age in months.
age_group	Breed-specific developmental age group.
barkunit_duration_seconds	Duration of the Bark Unit.
barkseq_duration_seconds	Duration of the parent Barkseq.
sub_bark_start_seconds	Bark Unit start time within the parent Barkseq.
sub_bark_end_seconds	Bark Unit end time within the parent Barkseq.
gmm_transcript_full	Full Elemental Dog Bark Unit transcript.
gmm_transcript_collapsed	Collapsed Elemental Dog Bark Unit transcript.

Original usernames, original video IDs, birthdates, source filenames, and local filesystem paths are not included in the public release.

Recommended Usage

The dataset provides two complementary representations:

• Bark Units (BUs): short isolated vocal elements extracted from Barkseqs.
• Barkseqs: parent vocalization segments from which Bark Units were extracted.

For self-supervised learning, representation learning, audio foundation model adaptation, and pretrained audio encoder training, we recommend using Barkseq audio whenever available. Barkseqs preserve more temporal context and are better suited for learning robust acoustic representations.

Bark Units are most appropriate for fine-grained EDBU analysis, discrete token modeling, vocal unit discovery, and Bark Unit-level classification.

Extracting the Shards

mkdir -p catvd_audio
cd catvd_audio

tar -xf ../shards/audio_barkunits_dogs_000001_000025.tar
tar -xf ../shards/audio_barkunits_dogs_000026_000050.tar
tar -xf ../shards/audio_barkunits_dogs_000051_000075.tar
tar -xf ../shards/audio_barkunits_dogs_000076_000100.tar
tar -xf ../shards/audio_barkunits_dogs_000101_000125.tar
tar -xf ../shards/audio_barkseqs_all.tar

After extraction, audio_barkunits/ and audio_barkseqs/ should be present.

Loading Metadata

from pathlib import Path
import pandas as pd

root = Path(".")
metadata = pd.read_csv(root / "barkunit_metadata.tsv", sep="	")

print(metadata.head())
print(len(metadata))

Loading Bark Unit Audio

from pathlib import Path
import pandas as pd
import soundfile as sf

root = Path(".")
metadata = pd.read_csv(root / "barkunit_metadata.tsv", sep="	")

row = metadata.iloc[0]
audio, sr = sf.read(root / row["barkunit_audio"])

print(audio.shape)
print(sr)

Loading Parent Barkseq Audio

from pathlib import Path
import pandas as pd
import soundfile as sf

root = Path(".")
metadata = pd.read_csv(root / "barkunit_metadata.tsv", sep="	")

row = metadata[metadata["has_barkseq_audio"] == 1].iloc[0]
audio, sr = sf.read(root / row["barkseq_audio"])

print(audio.shape)
print(sr)

Loading EDBU Transcripts

import pandas as pd

metadata = pd.read_csv("barkunit_metadata.tsv", sep="	")
row = metadata.iloc[0]

gmm_full = [int(x) for x in str(row["gmm_transcript_full"]).split()]
gmm_collapsed = [int(x) for x in str(row["gmm_transcript_collapsed"]).split()]

print(gmm_full)
print(gmm_collapsed)

Related Dataset

Researchers may also find our earlier DogSpeak dataset useful:

DogSpeak Dataset: https://huggingface.co/datasets/ArlingtonCL2/DogSpeak_Dataset

DogSpeak contains 77,202 bark sequences, referred to as Barkseqs, from 156 individual dogs across 5 breeds: Chihuahua, German Shepherd, Husky, Pitbull, and Shiba Inu. DogSpeak also provides the biological sex of the dog, labeled as male or female, associated with each vocalization.

Important Note on Dataset Overlap

Some individual dogs may overlap between DogSpeak and the Canine Age Transition Vocalization Dataset. If you use both datasets for training, evaluation, or benchmarking, please check for overlapping individuals and avoid train/test leakage. In many cases, it may be preferable to use one dataset or the other for a given experiment.

Citation

If you use this dataset in your research, please cite:

@inproceedings{lekhak2025data,
  title={A Data-driven Approach to the Longitudinal Study of Canine Vocal Pattern Development},
  author={Lekhak, Hridayesh and Dang, Tuan M. and Wang, Theron S. and Zhu, Kenny Q.},
  booktitle={Proceedings of the 33rd ACM International Conference on Multimedia},
  pages={12473--12482},
  year={2025},
  doi={10.1145/3746027.3758175}
}

License

The Canine Age Transition Vocalization Dataset is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. You may share and adapt the dataset for non-commercial purposes, provided that you give appropriate credit and distribute any derivative works under the same license.