KARAKALPAK SPEECH CORPUS: THE FIRST BENCHMARK DATASET FOR AUTOMATIC SPEECH RECOGNITION

While large-scale pre-trained models have significantly advanced multilingual Automatic Speech Recognition (ASR), many low-resource languages remain under-served due to the scarcity of high-quality annotated speech corpora. This paper introduces the Karakalpak Speech Corpus (KSC), the first publicly available benchmark dataset for Karakalpak, a Turkic language spoken by over two million people primarily in Karakalpakstan. The corpus encompasses 50 hours of predominantly read speech. The data was collected from 25 native speakers with a balanced gender distribution. To establish a performance benchmark, we fine-tuned the Wav2Vec 2.0 architecture, specifically evaluating the efficacy of transfer learning from multilingual pre-trained models.