UrduSpeech: 156-Hour Urdu Speech Corpus with Paralinguistic Metadata

156 hours of high-fidelity Urdu audio address a critical under-resourcing in speech technology. The corpus contains 71,792 diarized utterances across three specialized subsets: Standard Pakistani Urdu, Urdu-English Code-Switched, and Pakistani-Accented English. It was created by ASLP-lab and last updated in June 2026.

Use Cases

• Train automatic speech recognition models based on the large-scale, diarized Urdu audio.
• Develop code-switching detection algorithms based on the dedicated Urdu-English subset.
• Study paralinguistic features like emotion or speaker traits based on the 12-dimensional metadata.
• Build text-to-speech systems for Pakistani-accented English based on the US-EngPk subset.
• Benchmark speaker diarization performance on the 71,792 segmented utterances.

Strengths

• 156 hours of total audio provides substantial training material.
• 71,792 diarized utterances offer fine-grained segmentation.
• 12-dimensional paralinguistic metadata enables multifaceted analysis.
• Three distinct subsets (Standard, Code-Switched, English) support comparative studies.

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 update timestamp is from the future (2026).

Provenance

Source

ASLP-lab

Freshness

Last updated 2026-06-04 12:43:17.

Geography

Likely Pakistan, given the focus on Standard Pakistani Urdu and Pakistani-accented English.