Neyshekar is an open, community-driven Persian speech dataset collected via a web-based crowdsourcing platform at https://ney.shekar.io. It is designed to support research and development in text-to-speech (TTS), automatic speech recognition (ASR), speech representation learning, and other downstream Persian speech applications.
The recordings are provided by volunteer contributors, all of whom are native Persian speakers. Each release represents a stable snapshot of the dataset, enabling reproducible research and consistent benchmarking.
Neyshekar is released incrementally. Each release represents a stable snapshot of the dataset at the time of publication.
V6.0 — 2026-09-01 (download)
- Total samples: 62,279 (f: 33,735, m: 28,076)
- Informal samples: 15,222 (24.44%) (identified using Shekar rule-based InformalClassifier)
- Total duration (hours): 99.02
- Average clip duration (seconds): 5.72
- Total tokens: 701,621
- Vocab size: 29,535
- Speakers: 190
V6.0 is the first release to ship predefined splits. Splits are assigned per speaker, so a single recorder never straddles two sets and the evaluation splits stay speaker-disjoint from train. The speaker_id column makes this checkable: no identifier appears in more than one split.
| Split | Samples | Share | Hours | Speakers |
|---|---|---|---|---|
train |
58,244 | 93.5% | 91.99 | 134 |
validation |
1,886 | 3.0% | 3.14 | 26 |
test |
2,149 | 3.5% | 3.88 | 30 |
- Total samples: 50,026 (f: 26,826, m: 22,655)
- Total duration (hours): 79.22
- Average clip duration (seconds): 5.7
- Total tokens: 565,797
- Vocab size: 27,250
- Total samples: 40,008 (f: 21,613, m: 17,845)
- Total duration (hours): 63.03
- Average clip duration (seconds): 5.67
- Total tokens: 456,268
- Vocab size: 26,758
- Entities: 12,443 (identified using Shekar NER)
Caution
Sample mismatches in v4. A number of clips in the v4 release (2026-05-14) have misaligned audio–transcript pairs. Training or evaluating on this snapshot may introduce label noise and lead to unreliable results. If you have already downloaded v4, discard it and re-download v4.1.
- Total samples: 30019
- Total duration (hours): 45.71
- Average clip duration (seconds): 5.48
- Total tokens: 331714
- Vocab size: 23972
- Total samples: 20,020
- Total duration (hours): 29.08
- Average clip duration (seconds): 5.23
- Total tokens: 208,472
- Vocab size: 20,853
- Total samples: 10,044
- Total duration: 14.42 hours
- Average clip duration: 5.17 seconds
- Total tokens: 103,757
- Vocabulary size: 15,224
Any attempt to identify or uncover the identity of speakers in the Neyshekar datasets is strictly prohibited.
This dataset is released under the CC0 1.0 Universal license.
It may be used, modified, and redistributed for any purpose without restriction.
If you use Neyshekar in your research, please cite:
@article{amirivojdan2026neyshekar,
title={Neyshekar: An Open Persian Read-Speech Corpus for Automatic Speech Recognition},
author={Amirivojdan, Ahmad and Nadiri, Farzad and Alizadeh, Abolfazl and Yaraghi, Shaghayegh},
journal={arXiv preprint arXiv:2609.14542},
year={2026},
doi={10.48550/arXiv.2609.14542},
url={https://arxiv.org/abs/2609.14542}
}With ❤️ for IRAN