This paper presents a hybrid noise reduction algorithm for improving the quality of children's speech recorded in a real acoustic environment. The algorithm combines traditional spectral analysis methods such as fast Fourier transform (FFT) and Wiener filter with deep learning approaches, including recurrent neural networks (RNN) and long-term short-term memory (LSTM) networks. To evaluate the effectiveness of the proposed method, an audio corpus of children's speech in the Kazakh language was collected using the specially developed Telegram bot "Dataset Loader". The experimental part of the study showed a significant improvement in speech recognition performance after applying the noise reduction algorithm. In particular, the word recognition error rate (WER) decreased from 32.4% to 18.7%, and the F1-score increased from 0.72 to 0.88. Analysis of spectrograms showed a significant decrease in the background noise level and improved speech readability. The research results confirm the effectiveness of the hybrid approach for improving the quality of speech data and improving the accuracy of automatic speech recognition systems in complex acoustic conditions.