QASR: QCRI Aljazeera Speech Resource

QASR is the largest transcribed Arabic speech corpus with around 2,000 hours of data.
It features multi-layer annotation, covering multiple Arabic dialects and code-switching speech.

---

📘 Overview

QASR is a large-scale transcribed Arabic speech corpus collected from Aljazeera News Channel broadcasts.
The data is lightly supervised and linguistically segmented, designed to support a wide range of speech and language processing research tasks.

Key Features

• ~2,000 hours of transcribed Arabic speech
• Multi-dialect and code-switching coverage
• Multi-layer linguistic annotations
• Lightly supervised transcriptions
• Linguistically motivated segmentation

---

📄 Lisence

Non-Commercial Purpose ONLY!

---

📥 Download

You can request or download the dataset using the link below:

👉 Download QASR Dataset

Please follow the instructions on the linked page to complete the request process and download the data.

---

🧠 Applications

QASR is suitable for training and evaluating:

• Automatic Speech Recognition (ASR) systems
• Arabic Dialect Identification (acoustics- and linguistics-based)
• Punctuation Restoration
• Speaker Identification and Speaker Linking
• Spoken Language Understanding and other NLP modules for spoken data

---

📊 Data Source

The corpus was crawled from the Aljazeera news channel, providing rich diversity in topics, speakers, and dialectal variation.

---

📄 Citation

If you use QASR in your research, please cite:

@inproceedings{mubarak_qasr_2021,
  title     = {{QASR}: {QCRI} {Aljazeera} {Speech} {Resource}. {A} {Large} {Scale} {Annotated} {Arabic} {Speech} {Corpus}},
  booktitle = {{Proc. of ACL}},
  author    = {Mubarak, Hamdy and Hussein, Amir and Chowdhury, Shammur Absar and Ali, Ahmed},
  year      = {2021},
}