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

Source abstract

Building dependable automatic speech recognition (ASR) for Karakalpak is hindered by the limited amount of transcribed speech available for this language. Whisper handles a wide range of languages well, yet without adaptation its transcripts of Karakalpak remain unreliable, largely because the language combines agglutinative word formation with a Latin alphabet containing letters that do not occur in most other orthographies. Following the experimental protocol of recent studies on Whisper fine-tuning, we investigate how the model can be adapted using the Karakalpak Speech Corpus, an openly released speech-to-text collection that the authors assembled and checked. Three families of adaptation are compared within a single framework: updating the full model, updating selected layers only, and training lightweight modules added to an otherwise frozen network. The outcome is a reproducible reference point for Karakalpak ASR that may later serve applications in education, public administration, accessibility, and the digital preservation of the language.