Datasets:

disco-eth
/

WorldSpeech

Tasks:

Automatic Speech Recognition

Audio Classification

Modalities:

Languages:

Size:

ArXiv:

Tags:

DOI:

License:

Dataset card Files Files and versions

AAdonis commited on 17 days ago

Commit

b668db0

·

verified ·

1 Parent(s): a92b2e1

Update README.md

Files changed (1) hide show

README.md +1 -1

README.md CHANGED Viewed

@@ -676,7 +676,7 @@ configs:
 **Highly multilingual, large-scale paired speech corpus assembled from primary public archives such as national parliaments, public broadcasters, and public-domain audiobooks across 107 language-country configs, with strong non-European and low-resource coverage.**
-Every segment ships with a verified *human* transcript (not ASR pseudo-labels), an ASR transcript, the CER between them, a WADA-SNR estimate, and four DNSMOS-P.835 quality scores, so you can filter, rank, or stratify training data without rerunning a single signal-quality model.
 Audio is OGG Opus 24 kHz mono ~64 kbps VBR. Splits are 95/5, full provenance (`source`, `source_url`, `session_date`, `source_start_s`/`source_end_s`, `segment_id`) is preserved on every row, so you can trace any clip back to its original parliamentary record or broadcast.

 **Highly multilingual, large-scale paired speech corpus assembled from primary public archives such as national parliaments, public broadcasters, and public-domain audiobooks across 107 language-country configs, with strong non-European and low-resource coverage.**
+Every segment contains a verified *human* transcript, an ASR transcript, the CER between them, a WADA-SNR estimate, and four DNSMOS-P.835 quality scores.
 Audio is OGG Opus 24 kHz mono ~64 kbps VBR. Splits are 95/5, full provenance (`source`, `source_url`, `session_date`, `source_start_s`/`source_end_s`, `segment_id`) is preserved on every row, so you can trace any clip back to its original parliamentary record or broadcast.