Аннотация
This paper presents HybridKazASR, a hybrid automatic speech recognition (ASR) system for the Kazakh language that integrates three complementary acoustic models within a multi-stage pipeline. The system combines NVIDIA FastConformer (115M parameters, RNNT decoder) with two domain-adapted variants of Meta MMS-1B (965M parameters, CTC decoder): one fine-tuned on FLEURS read speech and another on the 559-hour Kazakh Speech Dataset (KSD). A key technical innovation is an efficient adapter-swap mechanism that switches between MMS variants by replacing only 138,348 parameters (~542 KB) in the CTC output layer, avoiding the need to load two full 3.9 GB models. The three hypotheses are fused via weighted 3-Way ROVER (Recognizer Output Voting Error Reduction) using Needleman-Wunsch alignment. The fused output undergoes two-stage language model rescoring: first with a 5-gram word-level KenLM, then with KazMorphLM – a novel morpheme-aware language model featuring a rule-based segmenter with 230 suffixes across 14 grammatical categories, a 7-gram morpheme model interpolated with a 5-gram word model at a 0.6/0.4 ratio using Witten-Bell smoothing, and a four-channel rescoring mechanism incorporating acoustic confidence, word-level LM, morpheme-level LM, and vowel harmony scores. Evaluated on the FLEURS test set (100 samples), HybridKazASR achieves a Word Error Rate (WER) of 6.30%, a normalized WER of 4.18%, and a Character Error Rate (CER) of 1.90%. Ablation studies confirm the contribution of each pipeline component, with KazMorphLM providing a 14.6% relative WER improvement over word-level KenLM rescoring alone.