Loading datasets...
Release Date: 11/6/2026
Format: FLAC, JSON, JSONL, MD, TSV
Size: 174.33 MB
The Pashto Single-Speaker Spontaneous Vlog Speech Corpus was developed by the Digital Pashto Association and PashtoAI as part of their work to build open speech resources for Pashto. The two organizations collected, transcribed, segmented and validated the material. The corpus contains 799 transcribed speech segments totaling 1.942 hours of natural, vlog-style Pashto speech. All recordings come from Muhib Vlog, a Pashto vlogger from central-southern Afghanistan who records himself in many different places as he travels and goes about daily life. The speech is unscripted and conversational: he explains what he sees, comments on events, and talks directly to his audience. His register moves between semi-formal and informal, and his delivery varies in tone and emotion, from calm explanation to excitement, humor and emphasis. Because the recordings were made in real-world settings, outdoors and in public places, they include natural background sounds and, at times, other voices in the background. This makes the corpus a realistic, challenging resource that reflects how Pashto is actually spoken in everyday video content, in contrast to clean studio or read speech. Each audio segment is paired with its Pashto transcript, produced by the Digital Pashto Association and PashtoAI team. The corpus is intended primarily for automatic speech recognition and speech-to-text, especially for building models that are robust to spontaneous speech, varied speaking styles and background noise. It is also useful for speech-text alignment, research on conversational and emotional speech, and evaluation of Pashto ASR on real-world audio. Muhib, the vlogger and main speaker, has given written consent for his recordings to be used and openly released in this dataset. The dataset is released under the CC0 1.0 Universal Public Domain Dedication, allowing anyone to use, copy, modify, redistribute and build on it for research, educational, commercial and other purposes, to the extent permitted by law. Because the corpus contains one main speaker and a single genre, it is not representative of all Pashto dialects, speakers or speaking styles. Background voices are incidental and are not the subject of the transcripts.
Released under the CC0 1.0 Universal Public Domain Dedication. The dataset owner confirms that Muhib, the vlogger and main speaker in the recordings, has given written consent for his recordings to be used and openly released in this dataset, including public redistribution, AI/ML training, research and commercial use. No additional dataset-specific restrictions are imposed.
You agree not to attempt to identify any person who may be audible in the background of the recordings, and not to use the recordings to create a synthetic voice imitating the speaker.
Ethical Review
The main speaker, Muhib, gave written consent for his recordings to be used and openly released in this dataset, including AI/ML use. Because the recordings were made in public places, other people may occasionally be audible in the background; they are not transcribed, and the dataset forbids attempts to identify them. No independent ethics-board certification is asserted.
Intended Use
Pashto automatic speech recognition and speech-to-text for spontaneous, real-world audio; noise-robust ASR; speech-text alignment; and research on conversational and emotional speech.
Composition: 799 clips totaling 1.942035 hours. Main speaker: Muhib, a male Pashto vlogger from central-southern Afghanistan. Source: Muhib Vlog recordings, made in many different locations. Developed by the Digital Pashto Association and PashtoAI, who collected, transcribed, segmented and validated the material. Domain: vlog-style spontaneous speech, semi-formal to informal, with varied tone and emotion.
Processing: Long recordings were fully decoded before segmentation. Audio was cut at source transcript timestamps rounded to sample boundaries and stored as 16-bit PCM FLAC. Original sample rates and channel configurations were preserved, with no resampling, denoising or gain adjustment, so real-world background sound is retained. Every FLAC file was decoded and compared sample-for-sample. Manifests (TSV and JSONL) retain original text, normalized text, source filenames, timestamps and hashes. Transcript normalization is limited to whitespace.
Selection: 2,260 source transcript entries = 1,455 inventory exclusions + 6 packaging exclusions + 799 retained clips. Originals and exclusion logs are preserved for provenance.
Limitations: Technical decoding confirms readability, not the accuracy of every transcript or segment boundary. Background noise and background voices are present in some clips; transcripts cover the main speaker only. No speaker-disjoint train/test split is provided.
