The models of Automatic Speech Recognition (ASR), like Whisper (OpenAI), attend a high accuracy on speech of adults but contend with children’s speech because of many factors such as articulation, spelling, speed. This study focuses on fine-tuning Whisper model on a small children's speech dataset to improve the accuracy of transcription. For data preprocessing, we started by extracting log-Mel spectrogram features from the audio data, adjusting their length and for an efficient training we tokenized the text labels. Using transfer learning, we fine-tune Whisper by adjusting some layers to better capture variations of the phonetic in children’s speech. We experiment with hyperparameters and training strategies to optimize the performance of the model. Evaluation metrics such as Word Error Rate (WER) reveal a 31.15% improvement in transcription accuracy compared to the baseline model. The fine-tuned model demonstrates greater robustness to pronunciation variability and short utterances, though challenges persist with rapid speech and background noise. Our results suggest that fine-tuning Whisper on children’s voice data significantly improves ASR performance, with applications in the education, accessibility tools, and healthcare. Future work includes expanding the dataset and refining the speech recognition model in multiple languages.