WorldSpeech

A multilingual ASR dataset containing over 65k hours of human transcribed speech across 127 language-region variants, drawn from national parliaments, public broadcasters, public-domain audiobooks, and international institutions. Rows consist of 24 kHz speech utterances paired with a human-provided transcript, an aligned ASR transcript, character error rate (CER) between the two, a WADA-SNR estimate, and four DNSMOS-P.835 quality scores.

Dataset Overview

How to use

from datasets import load_dataset

ds = load_dataset("disco-eth/WorldSpeech", "nl_nl", split="train")

row = ds[0]
wav    = row["audio"]["array"]         
sr     = row["audio"]["sampling_rate"]
text   = row["human_transcript"]
snr    = row["snr"]
dnsmos = row["dnsmos_ovr"]

Streaming

from datasets import load_dataset

ds = load_dataset("disco-eth/WorldSpeech", "nl_nl", split="train", streaming=True)

for row in ds:
    wav = row["audio"]["array"]
    break

Schema

Language Coverage

⚠️ Five configs (kh_km Khmer, la_lo Lao, mm_my Burmese, vn_vi Vietnamese, cn_ug Uyghur, sourced from Radio Free Asia include only metadata, transcripts, and source URLs. Audio must be downloaded separately from source_url.

Citation

If you use WorldSpeech in your research, please cite:

@misc{asonitis2026worldspeechmultilingualspeechcorpus,
      title={WorldSpeech: A Multilingual Speech Corpus from Around the World}, 
      author={Antonis Asonitis and Luca A. Lanzendörfer and Frédéric Berdoz and Roger Wattenhofer},
      year={2026},
      eprint={2605.09167},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.09167}, 
}

Dataset License

WorldSpeech is released under CC BY-NC 4.0. Individual source licenses apply per config and are listed in the Licenses section. Where a source license is more restrictive than CC BY-NC 4.0, the more restrictive terms govern that config.

Licenses