Skip to content
DSML.KZ
Publications

Article · Scientific Journal of Astana IT University · March 30, 2026

KAZMORPHLM: MORPHEME-AWARE LANGUAGE MODEL FOR KAZAKH AUTOMATIC SPEECH RECOGNITION

Yerlan Karabaliyev, Kateryna V. Kolesnikova

Kateryna Kolesnikova
Open via DOIPDF

Abstract

This paper presents KazMorphLM, a morpheme-aware language model for automatic speech recognition (ASR) in the Kazakh language. Kazakh belongs to the Turkic family and is characterised by a highly agglutinative morphology, in which a single root can generate a large number of inflected forms through productive suffixation. This property causes severe data sparsity for conventional word-level language models and reduces recognition accuracy. The proposed model introduces three main innovations. First, a rule-based morpheme segmenter uses an inventory of 230 suffixes across fourteen grammatical categories and includes phonological validation through vowel harmony and consonant assimilation rules. Second, a two-level interpolated n-gram architecture combines a 7-gram morpheme-level model with a 5-gram word-level model using an interpolation ratio of 0.6 to 0.4 and Witten–Bell smoothing. Third, a four-channel rescoring mechanism integrates acoustic confidence, word-level and morpheme-level language-model probabilities, and a vowel-harmony consistency score. KazMorphLM was integrated into a hybrid ASR pipeline combining NVIDIA FastConformer and Meta MMS-1B acoustic models. On the FLEURS test set, the system achieves a word error rate of 6.86%, a 14.6% relative improvement over word-level KenLM rescoring. The results indicate that higher-order morpheme modelling is essential for agglutinative languages and that corpus quality outweighs corpus size. The approach is applicable to other morphologically rich Turkic languages.

Speech Recognition and SynthesisNatural Language Processing TechniquesSpeech and Audio ProcessingKazakh languagemorphologically rich languagesagglutinative morphology

More by Kateryna Kolesnikova