Loading datasets...
License:
CC-BY-NC-4.0
Steward:
Audio Speech Language Processing lab(ASLP NPU)Dataset ID:
cmvl58msc01v807o52scz27un
Release Date: 10/23/2026
Format: .wav, .jsonl
Size: 71.87 GB
UrduSpeech is a large-scale, high-fidelity Urdu speech corpus comprising 156 hours of audio annotated with comprehensive 12-dimensional paralinguistic metadata. The corpus addresses the critical under-resourcing of Urdu in speech technology. To ensure standardized linguistic modeling, all dataset transcriptions are strictly normalized to the standard Right-to-Left (RTL) Urdu script, encompassing the complete standard alphabet: ا ب پ ت ٹ ث ج چ ح خ د ڈ ذ ر ڑ ز ژ س ش ص ض ط ظ ع غ ف ق ک گ ل م ن ں و ہ ھ ء ی ے
Licensing
Creative Commons Attribution Non Commercial 4.0 International (CC-BY-NC-4.0)
https://spdx.org/licenses/CC-BY-NC-4.0.htmlRestrictions/Special Constraints
By downloading or accessing the UrduSpeech dataset, users must agree to the following strict usage restrictions:
Non-Commercial Research Only: The dataset is released strictly for non-commercial research and academic purposes under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. Commercial deployment, monetization, or integration into paid products is strictly prohibited.
Prohibition of Malicious Use: The audio data and its derivatives must not be used for unauthorized voice cloning, deepfake generation, spoofing, harassment, deception, or any illegal activities.
No Unauthorized Redistribution: Users may not host, re-publish, or redistribute the raw dataset or its subsets to third parties without explicit prior acceptance and written permission from the dataset authors.
Data Provenance, Privacy, and Takedown Policy: All audio data is derived from openly available data and archival logs, and no personal identifiers or speaker identities are disclosed in the corpus. Users must not attempt to deanonymize or re-identify speakers. If a speaker requests it, their respective audio data will be promptly removed from the repository. Additionally, to prevent copyright infringement, no individual audio clip exceeds 35 seconds, and users must not attempt to stitch or reconstruct the original source broadcasts to bypass this limitation.
Ethical Compliance: Users must respect the data subjects and local cultural norms; the dataset is intentionally curated to be non-political and non-religious, and downstream models should not use this data to generate harmful or sensitive content.
Forbidden Usage
Ethical Review
The data collection and evaluation protocols for UrduSpeech adhere to established academic ethical standards and privacy frameworks: Informed Consent for Human Evaluators: For the human quality assessment and benchmark annotation experiments, all native-speaker evaluators participated in a controlled environment under written informed consent. Participants were fully briefed on the nature and scope of the research, the intended downstream applications, and how their evaluations would be reported. Work sessions adhered to ITU-T P.800 protocols, featuring isolated listening environments and mandatory rest breaks to prevent listener fatigue. Public Repository Provenance and Privacy Protection: Audio data was gathered exclusively from public repositories (YouTube and archival Pakistan Television logs). Processing is conducted under the legal basis of Legitimate Interest (Art. 6 GDPR) for non-commercial academic research to advance speech AI infrastructure for under-resourced languages. To enforce strict data minimization, all personal identifiers and metadata linked to real-world identities were stripped, residual noise and isolated-speaker clips were discarded, and speakers were assigned anonymized unique IDs across diarized clusters. Because individual notification of the 2,329 verified speakers requires a disproportionate effort, the dataset's public documentation serves as a transparency notice under the GDPR Article 14(5)(b) exception. Additionally, automated extraction methods respected domain-level technical opt-outs such as `robots.txt` where applicable. Content Filtering and Sensitive Data Safeguards: The dataset curation pipeline specifically filtered for non-political and non-religious content to prevent the capture or processing of sensitive personal beliefs, adhering strictly to local cultural norms. Copyright Compliance and Minimization: To prevent potential copyright infringement and avoid reconstructing long-form media, utterance extraction enforced an upper duration boundary of 35 seconds per clip. Right to Object and Takedown Framework: In line with data governance guidelines, a transparent takedown mechanism is maintained: any original speaker or copyright holder requesting the withdrawal of their voice data will have their respective audio segments promptly purged from the repository.
Intended Use
This dataset is intended for non-commercial research and development of Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) systems for the Urdu language. It serves as a foundational resource for training and evaluating speech models on complex linguistic phenomena, specifically targeting Right-to-Left (RTL) Perso-Arabic script constraints and pervasive Urdu-English code-switching. Furthermore, the inclusion of comprehensive 12-dimensional paralinguistic metadata allows researchers to utilize the corpus for specialized downstream speech AI applications, including speech emotion recognition, deepfake detection, speaker verification, and fine-grained acoustic analysis across diverse regional dialects and accents.
This dataset comprises two distinct legal components:
All text transcriptions, paralinguistic labels, and accompanying metadata produced by the authors are released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. You are permitted to share and adapt these materials for non-commercial purposes, provided appropriate attribution is given.
The underlying .wav speech files are sourced from publicly available, official Pakistani television channel broadcasts hosted on YouTube. The dataset creators claim no copyright or intellectual property over these original broadcasts. Audio distribution is conducted strictly under the Fair Dealing exception of Pakistan's Copyright Ordinance, 1962 (Section 57(1)(a)), which permits the use of copyrighted material solely for non-commercial "research or private study."
Statutory Basis for Fair Dealing: Pakistan Copyright Ordinance, 1962 (Act No. XXXIV of 1962), Chapter XII: Infringement of Copyright, Section 57(1)(a)(i) explicitly states:
"57. Certain acts not to be infringement of copyright. (1) The following acts shall not constitute an infringement of copyright, namely: (a) a fair dealing with a literary, dramatic, musical or artistic work for the purpose of— (i) research or private study;"
Legal Basis for Short-Clip Extraction :
To further ensure strict compliance and eliminate any residual commercial value, all musical compositions, background scores, and non-speech broadcast elements which carry highly sensitive, multi-layered copyrights were explicitly filtered out of this corpus.
We strictly respect the intellectual property and individual privacy rights of all parties:
If you use the UrduSpeech corpus or its metadata in your academic research or non-commercial projects, you are required to acknowledge our work by citing the following paper:
@misc{haq2026urduspeech156hoururduspeech,
title={UrduSpeech: A 156-Hour Urdu Speech Corpus with 12-Dimension Paralinguistic Annotations},
author={Attia Nafees ul Haq and Zeyu Zhu and Jingbin Hu and ChunJiang He and Lei Xie},
year={2026},
eprint={2605.17846},
archivePrefix={arXiv},
primaryClass={eess.AS},
url={[https://arxiv.org/abs/2605.17846](https://arxiv.org/abs/2605.17846)},
}
Compensated · similar languages
High-specularity optical edge-case dataset (425 assets) for benchmarking computer vision and autonomous navigation against severe specular reflections.
Versioned software-engineering corpus with code, configuration, workflows, and AI-tooling artifacts across frontend, backend, cloud, and infrastructure.