Adapting Whisper to Karakalpak: Full, Layer-Selective, and Parameter-Efficient Fine-Tuning for Low-Resource Speech Recognition

Building dependable automatic speech recognition (ASR) for Karakalpak is hindered by the scarcity of transcribed speech. Whisper handles many languages well, but without adaptation its Karakalpak transcripts remain unreliable, largely because the language combines agglutinative morphology with a Latin alphabet containing letters absent from most other orthographies. Following recent Whisper fine-tuning protocols, we adapt Whisper Large-v2 on the authors’ openly released Karakalpak Speech Corpus and compare full fine-tuning, layer-selective fine-tuning, and lightweight modules (bottleneck adapters and low-rank adaptation, LoRA) in an otherwise frozen network. On a speaker-independent test set, full fine-tuning reduces the word error rate (WER) from 37.70% to 13.53% and the character error rate from 15.40% to 5.42%. Freezing the lower 13 encoder layers gives comparable accuracy (13.39% WER), and adapters in the upper encoder layers and the decoder reach 15.59% WER with only 5.9 M trainable parameters. Substituted special letters remain the most frequent error type. Full fine-tuning raises WER on closely related Kazakh and Uzbek by 4.6 and 6.6 percentage points, whereas adapters and LoRA largely preserve multilingual accuracy. The results provide a reproducible reference point for Karakalpak ASR that may later serve applications in education, public administration, accessibility, and digital language preservation.