Low-resource Dialect Speech Recognition: a Case Study of Adapting Whisper to the Khorezm Dialect of Uzbek
Year: 2026
Type: article
Abstract: The Khorezm dialect of Uzbek differs from the literary standard in phonology and vocabulary and retains Oghuz features, yet no speech recognition system has been reported for it. The purpose of this study is to build and evaluate a model that transcribes Khorezm dialect speech in its dialectal form. A 4.65-hour corpus of manually transcribed dialect speech was prepared, and the Whisper-large-v3-turbo model was adapted with low-rank adaptation (LoRA), training 3.33% of its parameters. In a second stage, 18.61 hours of untranscribed Telegram voice messages were transcribed by the adapted model; after rule-based and confidence-based filtering, 13.24 hours of pseudo-labelled speech were added to the training data. Without adaptation, Whisper reached a word error rate (WER) of 86.64% and a character error rate (CER) of 29.67% and often wrote dialect speech in Turkish, Azerbaijani or Kazakh orthography. LoRA adaptation reduced WER to 32.23% and CER to 7.86%. Pseudo-labelling gave 31.84% and 7.81% on read speech, a small difference that was not tested for significance; on a 25-word sample of spontaneous speech WER fell from 68.0% to 48.0%, which is a preliminary observation. A few hours of dialect data are therefore sufficient for a usable baseline, whereas the benefit of untranscribed social-media speech for conversational dialect remains to be confirmed on a larger spontaneous test set.
Author: Ruzikajon Kurbonova
Institution: Urgench State University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225997581
