Vowel-Harmony Constraints in Kazakh ASR: A Systematic Investigation of Integration Approaches
Year: 2026
Type: article
Abstract: Kazakh vowel harmony is well established linguistically, but its explicit integration into modern neural automatic speech recognition (ASR) remains insufficiently studied. This work investigates whether a productive front/back vowel-harmony constraint can be operationalized within Whisper large-v3 for Kazakh ASR. Two complementary strategies are evaluated on the ISSAI Kazakh Speech Corpus 2: (1) a differentiable token-level harmony regularizer added during fine-tuning and (2) harmony-based N-best hypothesis re-ranking during inference. We further introduce the Harmony Violation Rate (HVR) as a diagnostic measure and analyze Whisper’s byte-level BPE representation of Kazakh vowels. Neither integration strategy produces a statistically significant improvement in test WER. Re-ranking yields a small but statistically significant reduction in HVR, while additional analyses show no recognition disadvantage for byte-fragmented vowels and no systematic reduction in front/back vowel confusions after harmony-constrained fine-tuning. These findings show that simple token-level regularization and sentence-level re-ranking do not adequately model the contextual stem–suffix dependency underlying Kazakh vowel harmony. The study therefore identifies specific limitations of straightforward phonological-constraint integration and provides diagnostic evidence for future representation- and morphology-aware Kazakh ASR.
Source: Applied Sciences
DOI: 10.3390/app162111035
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7225634333