Fine-Tuning Whisper for Low-Resource Automatic Speech Recognition in the Karakalpak Language

Abstract The Karakalpak language is still poorly represented in supervised speech resources, making reliable automatic speech recognition (ASR) difficult to build. Whisper performs well in many multilingual settings, but its unadapted output is not yet dependable for Karakalpak, where orthography, agglutinative morphology, and language-specific Latin characters require careful handling. This study adapts the experimental design of recent Whisper fine-tuning work to the Karakalpak Speech Corpus, a public speech-to-text resource constructed and verified by the authors. We evaluate a Karakalpak ASR framework that includes full-model fine-tuning, layer-selective fine-tuning, and parameter-efficient adaptation through additional trainable modules. The results provide a reproducible baseline for Karakalpak speech recognition and support future speech technologies for education, public administration, accessibility, and digital language preservation.