LakoreAI/clotho-dev-sample

Clotho Development Subset

A sampled subset (~3GB) of the Clotho v2.1 development split, packaged for quick experimentation with audio-text retrieval pipelines.

📋 Dataset Description

This dataset is a convenience subset of the Clotho audio captioning dataset, created for rapid prototyping and testing of audio-text retrieval models (e.g., CLAP fine-tuning) on limited compute.

• —Source: Clotho v2.1 (development split)
• —Original Authors: K. Drossos, S. Lipping, T. Virtanen
• —Original Paper: Clotho: An Audio Captioning Dataset

📊 Dataset Structure

Splits

Features

Audio Details

• —Duration: 15–30 seconds per clip
• —Sample Rate: 44,100 Hz
• —Channels: Mono
• —Format: WAV (stored as Parquet/Arrow on Hub)

🚀 Usage

from datasets import load_dataset

ds = load_dataset("your-username/clotho-dev-sample")

# Access a sample
sample = ds["train"][0]
print(sample["caption_1"])   # "A dog barks in the distance"
print(sample["audio"])       # {'array': array([...]), 'sampling_rate': 44100}

With CLAP

from transformers import ClapProcessor, ClapModel

processor = ClapProcessor.from_pretrained("laion/clap-htsat-unfused")
model = ClapModel.from_pretrained("laion/clap-htsat-unfused")

sample = ds["train"][0]
inputs = processor(
    audios=sample["audio"]["array"],
    sampling_rate=sample["audio"]["sampling_rate"],
    text=sample["caption_1"],
    return_tensors="pt",
    padding=True,
)
outputs = model(**inputs)

⚠️ Important Notes

• —This is a subset (~43%) of the full Clotho development split, sampled randomly with seed=42
• —For official benchmarking, use the full Clotho dataset from Zenodo
• —This subset is intended for pipeline testing and prototyping only

📄 Citation

If you use this dataset, please cite the original Clotho paper:

@inproceedings{drossos2020clotho,
  title={Clotho: An Audio Captioning Dataset},
  author={Drossos, Konstantinos and Lipping, Samuel and Virtanen, Tuomas},
  booktitle={ICASSP 2020 - IEEE International Conference on Acoustics, Speech and Signal Processing},
  pages={736--740},
  year={2020},
  organization={IEEE}
}

🏷️ License

This dataset follows the original Clotho license. Audio clips are sourced from Freesound under Creative Commons licenses. Please refer to the original dataset for full license details.