Abstract
This study examines the challenge of automatically transcribing multilingual institutional speech in Kazakhstan, where speakers frequentlбукy switch between Kazakh, Russian, and English. While modern automatic speech recognition (ASR) systems achieve high accuracy in monolingual settings, their performance often declines in speech containing intensive language interaction and morphological mixing. In addition, the limited availability of annotated multilingual data reduces the effectiveness of retraining-based methods. To address this issue, a multilingual transcription system is proposed that improves ASR output without altering acoustic models. The system combines segmentation refinement, token-level language identification, contextual tagging, morphological analysis, normalization, and N-best hypothesis rescoring. It is implemented on the Whisper Large-v2 baseline and evaluated under inter-sentential, intra-sentential, and intra-word code-switching conditions. The experimental results show consistent improvements across all evaluation scenarios. These findings indicate that the proposed approach provides an effective and computationally efficient solution for multilingual meeting transcription. The system also demonstrates strong practical potential for institutional documentation in linguistically diverse environments.