Аннотация
This paper considers the problem of punctuation restoration in Kazakh automatic speech recognition (ASR) systems. The proposed approach demonstrates a solution to the punctuation restoration problem in Kazakh ASR for four main punctuation classes (COMMA, PERIOD, QUESTION, and EXCLAMATION) and the unpunctuated class (O). Experiments were conducted on both monolingual Kazakh texts and mixed Kazakh-Russian data, covering the practical parameters of multilingual communication in modern ASR applications. To reduce the imbalance between the exclamation and question mark classes, a specially labeled corpus was created using both real sentences collected from open Kazakh-language sources and LLM-assisted paraphrasing. Experiments showed that the punctuation restoration model achieved a macro-F1 score of 93.23% for four punctuation classes (COMMA, PERIOD, QUESTION, and EXCLAMATION) on the monolingual test set, with F1 scores of 98.99% for period, 94.35% for question mark, 87.77% for comma, and 91.82% for exclamation mark. On a held-out code-switching test set, adding code-switched sentences to training raised macro-F1 from 32.81% to 63.88%, though the exclamation class remains the weakest (26.47% F1), pointing to prosody-free intonation cues as the main remaining challenge.