Zainab Sorani Kurdish Speech Corpus: A Transcribed Audio Dataset for Automatic Speech Recognition

Description

This repository holds the Zainab Sorani Kurdish Speech Corpus, a large collection of audio recordings that have been manually transcribed word by word. The dataset was built to help researchers develop and improve Automatic Speech Recognition (ASR) systems for Sorani Kurdish — a dialect that has been largely overlooked in the world of language technology, despite being spoken by tens of millions of people. Kurdish as a whole is spoken by an estimated 30 to 40 million people across Iraq, Iran, Turkey, and Syria, and Sorani is particularly dominant in the Kurdistan Region of Iraq and parts of Iran. Rather than recording speakers in a studio or lab, the audio was taken from real Kurdish broadcasts — podcasts and TV discussion and meeting programmes. This was a deliberate decision, because real broadcast speech is far more representative of how people actually talk: it includes natural pauses, varied speaking styles, and a wide range of topics. Before anything was transcribed, all recordings went through a careful music-removal stage to clean out background music and any non-speech sounds, making sure that the spoken words remained completely clear and undistorted. Everything that followed — segmentation, transcription, and checking — was done by hand. The end result is a clean, reliable corpus of 12,367 audio clips covering around 28.89 hours of Sorani Kurdish speech.

Institutions

• Duhok Polytechnic University

  Duhok, Dihok
• Akre University for Applied Sciences

  Nineveh, Akre

Categories

Licence