atikuwu/karakalpak-speech-corpus

📚 Karakalpak Speech Corpus (107 Hours)

The Karakalpak Speech Corpus is the first comprehensive, open-access, community-crowdsourced speech recognition dataset for the Karakalpak language (kaa), a low-resource Turkic language spoken primarily in the Republic of Karakalpakstan (Uzbekistan).

Founded and led by Atabek Kadirbergenov alongside a student research team from the Muhammad al-Khwarizmi Specialized School in Nukus, this dataset was created to preserve cultural heritage, bridge the digital divide, and empower modern AI technology for the Karakalpak language.

---

🎯 Project Mission

Digital extinction is a growing threat to minority languages. Without open speech corpora, state-of-the-art voice recognition, transcription engines, and conversational AI cannot support native speakers. This 107-hour corpus provides the community with an open scientific foundation for speech technology.

---

👥 Project Leadership & Team (Muhammad al-Khwarizmi Specialized School, Nukus)

• Atabek Kadirbergenov — Founder, Project Lead & AI/ML Engineer
  Responsibilities: Conceived and founded the project, assembled and led the team, curated the speech dataset, developed phonetic text normalization algorithms, implemented acoustic noise filtering, designed stratified evaluation splits, and benchmarked Whisper ASR models.
  Telegram: @atik_uwu

• Sanjar Tleumuratov — Telegram Bot, Systems & Data Moderation
  Responsibilities: Crowdsourcing bot architecture, backend server infrastructure, contributor administration, submission quality moderation (filtering invalid/spam recordings), and audio ingestion pipeline.
  Telegram: @Sanjar030609 | LinkedIn

• Dawitbay Nasiratdinov — Media, Outreach & PR Lead
  Responsibilities: Regional public relations, community engagement, promotional video production, editing, and mobilizing 200+ native contributors.
  Telegram: @Dawitbay_Nasiratdinov | LinkedIn

---

🙏 Acknowledgements

We express our deepest gratitude to:

• Muhammad al-Khwarizmi Specialized School (Nukus): For granting GPU computing infrastructure to train the Whisper Medium model and actively supporting our initiative across the academic community (Official School Announcement).
• Zulfiya Erniyazova (@erniyazovaa_a): For her significant individual contribution in recording multiple hours of high-quality voice data for this dataset.
• The Karakalpak Community: All 200+ contributors across Karakalpakstan who recorded and donated their voices.

---

📈 Dataset Overview

• Total Duration: 106.92 hours
• Total Utterances: 21,702 verified audio samples
• Speakers: 200+ unique contributors
• Audio Format: OGG (Opus/Vorbis), 16 kHz, single-channel mono
• Orthography: Karakalpak Latin script (á, ǵ, ı, ń, ó, ú, etc.)
• Text Normalization: Numerals and abbreviations expanded into complete words (100 -> júz, 70 -> jetpis).

---

📂 Dataset Splits

Partitioned using speaker-stratified sampling to guarantee speaker independence between training and test sets:

Split	Samples	Duration	Purpose
Train	19,554	~96.3 hrs	Primary model training
Validation	1,074	~5.3 hrs	Hyperparameter tuning & validation
Test	1,074	~5.3 hrs	Independent benchmark evaluation
Total	21,702	106.92 hrs

---

📋 Data Format & Usage (Apache Parquet)

The dataset is packaged in Apache Parquet (.parquet)

Python Quickstart:

from datasets import load_dataset

# Load full dataset (train, validation, test)
dataset = load_dataset("atikuwu/whisper-medium-karakalpak")

# Inspect a sample
sample = dataset["train"][0]
print("Transcription:", sample["text"])
# Audio waveform is automatically decoded:
# sample["audio"]["array"], sample["audio"]["sampling_rate"]

# Stream dataset on-the-fly (without downloading all 107 hours)
streamed_ds = load_dataset("atikuwu/whisper-medium-karakalpak", streaming=True)
for item in streamed_ds["train"].take(5):
    print(item["text"])

Schema & Fields:

• id (int64): Unique anonymized utterance index.
• audio (Audio feature): Native audio dictionary containing embedded binary bytes, sampling rate (16 kHz), and file path.
• text (string): Normalized Karakalpak text in Latin orthography with numerals expanded into words.
• raw_text (string): Unnormalized source text as originally presented to the speaker.

---

🔒 Privacy & Full Anonymization

• Zero PII: To protect contributor privacy, all speaker identifiers, Telegram usernames, user IDs, and local machine directory paths have been completely purged from the public release.
• Consent: All audio recordings were contributed voluntarily under explicit terms authorizing open scientific and public dissemination under the CC-BY-4.0 license.

---

📜 Citation

@dataset{kadirbergenov2026karakalpakdataset,
  title={Karakalpak Speech Corpus: A 100-Hour Crowdsourced Speech Dataset for Karakalpak ASR},
  author={Kadirbergenov, Atabek and Tleumuratov, Sanjar and Nasiratdinov, Dawitbay},
  organization={Muhammad al-Khwarizmi Specialized School, Nukus},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/atikuwu/whisper-medium-karakalpak}}
}