WorldSpeech: 65,000 Hours of Multilingual Speech from Parliaments and Broadcasters

Over 65,000 hours of human-transcribed speech across 127 language-region variants form this multilingual ASR dataset. The data is drawn from sources including national parliaments, public broadcasters, audiobooks, and international institutions. Each row consists of a 24 kHz speech utterance paired with a human transcript, an aligned ASR transcript, a character error rate, a WADA-SNR estimate, and four DNSMOS-P.835 quality scores.

Use Cases

• Training multilingual ASR models based on the 127 language-region variants.
• Benchmarking ASR system performance using the provided human and aligned ASR transcripts.
• Analyzing speech quality and noise levels using the WADA-SNR and DNSMOS-P.835 quality scores.
• Evaluating transcription accuracy by comparing the character error rate (CER) between human and machine transcripts.
• Studying formal speech patterns from sources like parliamentary proceedings and public broadcasts.

Strengths

• Contains over 65,000 hours of transcribed speech, indicating a large-scale resource.
• Covers 127 language-region variants, suggesting broad linguistic diversity.
• Includes multiple quality metrics per utterance: CER, WADA-SNR, and four DNSMOS-P.835 scores.
• Sources data from authoritative public-domain institutions like parliaments and broadcasters.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Freshness should be verified as the last metadata update was 2026-05-17.

Provenance

Source

National parliaments, public broadcasters, public-domain audiobooks, and international institutions.

Collection Method

Likely aggregated from publicly available audio sources and transcribed.

Freshness

Last updated 2026-05-17 17:16:31; freshness should be verified.

Geography

Global, implied by the 127 language-region variants and international sources.