This study aims to improve automatic speech recognition for children aged 5–12 by accounting for the acoustic and phonetic characteristics of the Kazakh language. The objective is to adapt OpenAI’s Whisper-tiny architecture to Kazakh through fine-tuning on a specialized children’s speech dataset. A two-stage fine-tuning strategy was applied: the model was first trained on 235 audio samples and then on an expanded corpus of 1,936 samples. The audio recordings were standardized to a sampling rate of 16 kHz, and the text transcriptions were normalized. Model performance was evaluated using Word Error Rate (WER) and Accuracy. The experimental results showed that WER decreased from 0.80 to 0.42, representing a relative improvement of 47.5%, while Accuracy increased from 20.43% to 57.95%. These findings demonstrate that increasing the amount of data and applying domain-specific fine-tuning improve the recognition of children’s speech. The proposed approach can support the development of educational and intelligent speech systems for low-resource languages.