SHALA KAZAKH: A MIXED TRANSCRIPTION OF KAZAKH AND RUSSIAN

Abstract

This thesis addresses the challenge of Automatic Speech Recognition (ASR) for "Shala Kazakh", a widespread code-switching phenomenon in Kazakhstan, where Kazakh speakers integrate Russian words and expressions in their speech. While ASR systems are rapidly improving for high-resource languages, they struggle to recognise code-switching scenarios, especially in low-resource languages, like Kazakh. This research introduces a novel approach by training a state-of-the-art (SOTA) Whisper model on both monolingual Kazakh and Russian datasets, additionally training it on a freshly collected 52-hour code-switching dataset that captures bilingual speech patterns gathered from TikTok. The experimental results demonstrate that incorporating a Russian dataset significantly improves transcription for the code-switching scenario. This work provides a framework for developing robust ASR systems for other low-resource languages like Kazakh with similar code-switching scenarios, contributing both technological advances and language preservation.