Neyshekar
A Large-Scale Open Persian Speech Dataset
Neyshekar is an open, community-driven Persian speech dataset collected via a web-based crowdsourcing platform at ney.shekar.io. Recordings are provided by volunteer contributors and paid voice actors, all native Persian speakers. Each release is a stable snapshot enabling reproducible research and consistent benchmarking. Released entirely under CC0 1.0 Universal. Free for any use, including commercial.
Dataset Statistics: v6.0 (Latest)
Speaker gender distribution: 33,735 female recordings and 28,076 male recordings.
Informal-register samples are identified using the Shekar rule-based
InformalClassifier, so colloquial and formal speech can be filtered or balanced during training.
Predefined Splits
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. This makes results directly comparable across papers without each team inventing its own partition.
Named Entity Distribution
Entities were identified automatically using the Shekar NER model, providing rich linguistic metadata for downstream tasks. The counts below are measured on the v4.1 snapshot (12,443 entities in total).
Why Neyshekar?
Open & CC0 Licensed
Released under CC0 1.0 Universal, the most permissive open license. Use it for any purpose, including commercial products and proprietary models, without attribution requirements.
Native Speakers Only
All recordings are from native Persian speakers: a mix of volunteer contributors and professional voice actors, ensuring natural prosody and authentic pronunciation.
Speaker-Disjoint Splits
Since v6.0, every release ships predefined train, validation, and test splits assigned per speaker, so no recorder appears in more than one set and benchmark numbers reflect true generalisation.
Reproducible Releases
Each version is a stable, versioned snapshot on Zenodo with a DOI, enabling reproducible experiments and consistent benchmarks across publications.
Rich Metadata
Every release includes transcriptions, speaker gender, duration statistics, vocabulary counts, register labels, and NER-tagged entity annotations from the Shekar NLP pipeline.
Community-Driven Growth
The dataset grows continuously through community contributions at ney.shekar.io. Each new release captures the latest snapshot with incremental improvements.
Balanced Gender Coverage
v6.0 includes 33,735 female and 28,076 male recordings, a balanced split essential for building gender-robust speech models.
Formal & Informal Speech
About 24% of v6.0 (15,222 samples) is colloquial Persian, labelled by the Shekar InformalClassifier, so models can be trained or evaluated on everyday spoken register rather than formal reading style alone.
Use Cases
Neyshekar is designed for a broad range of Persian speech research and engineering tasks:
Version History
Neyshekar is released incrementally. Each version is archived on Zenodo with a permanent DOI.
Note on v4. A number of clips in the v4 release (May 14, 2026) have misaligned audio–transcript pairs. Training or evaluating on that snapshot may introduce label noise and lead to unreliable results. If you have already downloaded v4, discard it and re-download v4.1 or later.
License & Terms of Use
CC0 1.0 Universal
The Neyshekar dataset is released under the CC0 1.0 Universal (Public Domain Dedication) license. It may be used, modified, and redistributed for any purpose, including commercial use, without restriction or attribution.
Note: Any attempt to identify or uncover the identity of individual speakers in the dataset is strictly prohibited.
Frequently Asked Questions
What is Neyshekar?
Neyshekar is a large-scale open Persian speech dataset built through community crowdsourcing. It contains 62,000+ recordings totalling 99+ hours of native Persian speech, along with transcriptions, gender labels, formal/informal register labels, NER-tagged entity metadata, and predefined speaker-disjoint splits.
How many hours of speech does Neyshekar contain?
Version 6.0 (the latest, released September 2026) contains 99.02 hours of audio across 62,279 recordings averaging 5.72 seconds each. The dataset has grown from 14.42 hours in v1 to 99+ hours in v6.0.
Does Neyshekar include informal Persian speech?
Yes. v6.0 contains 15,222 informal-register samples (24.44% of the dataset), labelled using the Shekar rule-based InformalClassifier. This supports training and evaluating models on colloquial spoken Persian rather than formal reading style alone.
Which version should I use?
Use v6.0, the latest release, which also ships the predefined splits. Avoid v4: it contains misaligned audio–transcript pairs and was superseded by v4.1. If you downloaded v4, discard it and re-download.
Where can I download the Neyshekar dataset?
Download it from Zenodo: doi.org/10.5281/zenodo.18073632. The collection platform is at ney.shekar.io.
Can I use Neyshekar for commercial products?
Yes. The CC0 1.0 Universal license places the dataset in the public domain. You can use, modify, and redistribute it for any purpose, including building commercial ASR or TTS products, without restriction.
Who recorded the dataset?
Recordings were contributed by volunteer community members at ney.shekar.io and by paid professional voice actors. All speakers are native Persian speakers.
What tasks is Neyshekar designed for?
Neyshekar targets ASR, TTS, speech representation learning, speaker identification, voice activity detection, and other downstream Persian speech applications.
Does Neyshekar provide train/validation/test splits?
Yes, since v6.0. The release ships predefined splits — train (58,244 samples, 91.99 h), validation (1,886 samples, 3.14 h), and test (2,149 samples, 3.88 h). Splits are assigned per speaker, so a recorder never straddles two sets and the evaluation splits are speaker-disjoint from train.
How do I cite Neyshekar?
Cite the Zenodo record: Amirivojdan, A. (2026). Neyshekar: A Large-Scale Open Persian Speech Dataset (v6.0). Zenodo. DOI: 10.5281/zenodo.18073632.
Citation
If you use Neyshekar in your research, please cite:
@dataset{amirivojdan_neyshekar_2026,
author = {Amirivojdan, Ahmad},
title = {{Neyshekar: A Large-Scale Open Persian Speech Dataset}},
year = {2026},
version = {v6.0},
publisher = {Zenodo},
doi = {10.5281/zenodo.18073632},
url = {https://doi.org/10.5281/zenodo.18073632}
}