Loading datasets...
Release Date: 10/20/2026
Format: FLAC, TSV, JSONL, JSON, MD
Size: 782.08 MB
The Pashto Multi-Speaker Informational Speech Corpus, prepared by the Digital Pashto Association and PashtoAI, contains 3,880 transcribed Pashto speech clips totaling 8.914 hours. The corpus is built from publicly presented informational video content produced by five male Pashto speakers, with written permission obtained from the contributors for use of their audio and video materials. The contributors represent several major Pashto-speaking regions. Abdul Hai Kakar is from Zhob in South Pakhtunkhwa, Abid Haiderzai is from Kabul, Noorul Bari Mal is from Kandahar, Shahkar Utmankhel is from North Pakhtunkhwa, and Ihsan Kamal is from Paktia in central/eastern Afghanistan. All contributors are male and approximately 25–45 years old. The source material consists primarily of YouTube-style informational presentations in which the speakers explain topics, share knowledge, or address their audiences directly. Because the contributors use a broadly similar presentation style while representing different geographic and linguistic backgrounds, the corpus provides useful variation in pronunciation, accent, speaking style, and regional Pashto within a relatively consistent content format. The source videos were collected with written permission, transcribed into Pashto, and subsequently reviewed against the corresponding audio. The resulting dataset contains FLAC audio with aligned transcript files, source and speaker metadata, and segment-level provenance information. The corpus is particularly suitable for Pashto text-to-speech and speech-synthesis research because it contains clear, presentation-style speech from multiple speakers with relatively good recording quality. It can also support automatic speech recognition, speech-to-text systems, speech-text alignment, pronunciation and phonetic research, accent and regional speech analysis, acoustic modeling, speaker-aware modeling, and other Pashto speech and language technologies. The combination of speakers from Afghanistan and different parts of Pakhtunkhwa provides broader regional coverage than a single-speaker corpus while maintaining a relatively consistent informational speaking style. This makes the dataset useful for developing Pashto speech systems that need both speaker diversity and reasonably consistent speech characteristics. The dataset owner confirms that signed written permissions have been obtained from all five contributors—Abdul Hai Kakar, Abid Haiderzai, Noorul Bari Mal, Shahkar Utmankhel, and Ihsan Kamal—for public redistribution, AI and model-training use, and release under the selected CC0 1.0 Universal Public Domain Dedication, to the extent of the rights they hold. Most source materials were published on the contributors’ own YouTube channels; the release authorization is based on their signed permissions.
Restrictions/Special Constraints
The dataset owner confirms that signed written permissions have been obtained from all five contributors—Abdul Hai Kakar, Abid Haiderzai, Noorul Bari Mal, Shahkar Utmankhel, and Ihsan Kamal—for public redistribution, AI and model-training use, and release under the selected CC0 1.0 Universal Public Domain Dedication, to the extent of the rights they hold. Most source materials were published on the contributors’ own YouTube channels; the release authorization is based on their signed permissions. No additional dataset-specific reuse restrictions are imposed on rights covered by the CC0 dedication. Rights in any third-party material outside the contributors’ control are not granted by this statement.
Forbidden Usage
No additional dataset-specific prohibitions are imposed on rights covered by the selected CC0 dedication. Applicable law and any third-party rights remain unaffected.
Ethical Review
The dataset owner confirms signed written permission from all five named contributors for public redistribution, AI/model-training use, and the selected CC0 release within their rights. No independent ethics certification is asserted. Technical processing checks do not constitute a comprehensive privacy, third-party-voice, or sensitive-content review.
Intended Use
Pashto text-to-speech and speech synthesis, automatic speech recognition, speech-text alignment, pronunciation and phonetic research, acoustic modeling, speaker-aware modeling, language modeling, and other AI and speech-language research under the selected CC0 dedication, to the extent of the rights held by the contributors and dataset creators.
The corpus contains 3,880 speech clips totaling 8.914 hours from five male Pashto speakers aged approximately 25–45. The contributors represent different Pashto-speaking regions, including Paktia, Kabul, Kandahar, Zhob in South Pakhtunkhwa, and North Pakhtunkhwa. All source materials were used with written permission from the contributors.
The source content consists mainly of informational and explanatory video presentations originally produced for online audiences. Although the speakers discuss different subjects, their presentation style is broadly similar: clear, direct, continuous speech intended to explain information to viewers. This provides a useful balance between regional speaker diversity and relatively consistent speaking style.
Source videos were transcribed into Pashto and the transcripts were subsequently reviewed against the corresponding audio. The final package contains segmented FLAC audio, transcript manifests, speaker and source metadata, and provenance information linking each segment to its original source.
The corpus is particularly suitable for text-to-speech, speech synthesis, automatic speech recognition, speech-to-text, acoustic modeling, pronunciation research, speaker-aware modeling, and regional Pashto speech analysis. The relatively clear recording quality and presentation-style delivery may also make the data useful for developing natural-sounding Pashto synthetic voices after appropriate model-specific preparation.
The corpus contains multiple speakers and regional varieties but should not be treated as fully representative of all Pashto dialects, genders, age groups, or speaking styles. All contributors are male and fall within a relatively narrow adult age range. Researchers should also consider speaker-specific train, validation, and test splits when evaluating models to avoid speaker leakage.
The accompanying metadata and provenance records should be consulted for speaker information, source mapping, transcription details, and processing history.
