aranemini/southern-kurdish-asr

Bestun: Southern Kurdish speech recognition resources and benchmarking

This repository provides speech recognition (ASR) resources for Southern Kurdish (ISO 639-3: sdh), a threatened variant of the Kurdish macrolanguage.
It includes:

• Bestun training corpus: ~30 hours of manually validated read speech
• Evaluation benchmark: 773 validated utterances (86.74 minutes) recorded by 8 speakers from different Southern Kurdish vernacular regions

The dataset and models are released under CC BY-NC-ND 4.0.

---

Contents and splits

The dataset is organized into two main splits:

• train: Bestun training corpus (crowdsourced read speech)
• test: Southern Kurdish evaluation benchmark (100 sentences × multiple speakers; validated subset)

Recording environments (training)

Training data were collected with two crowdsourcing tools:

• telegram_bot: recorded via a Telegram bot
• web_tool: recorded via a web-based recording tool jiridastkrd.com

---

Data statistics (from the paper)

Training corpus (Bestun)

• Audio format: mono, 24 kHz, WAV
• Total recordings: 18,636
• Unique utterances: 15,273
• Duration: ~30 hours
• Speakers: 208 (age range 18–49)

Evaluation benchmark

• Speakers: 8 (from different vernacular regions)
• Validated utterances: 773
• Duration: 86.74 minutes

---

Files and columns

Each example contains an audio field and metadata. Typical columns:

• audio: speech waveform (WAV)
• kurdish: transcription in Southern Kurdish Arabic-based script
• speaker_id: anonymized speaker identifier
• gender: speaker gender when available (M / F)
• region: vernacular / region label when available
• subset: optional source label used during dataset construction

---

How to download / use

Option 1 — Use datasets in Python

from datasets import load_dataset

ds = load_dataset("aranemini/southern-kurdish-asr")
print(ds)
print(ds["train"][0])

Citation

If you use this dataset, please cite our paper:

@inproceedings{mohammadamini2026southernkurdish,
  title     = {Southern Kurdish speech recognition resources and benchmarking},
  author    = {Mohammad Mohammadamini and Marie Tahon},
  booktitle = {Proceedings of the 2026 Language Resources and Evaluation Conference (LREC 2026)},
  year      = {2026},
  address   = {Palma, Mallorca (Spain)},
  publisher = {European Language Resources Association (ELRA)},
}