This research explores the classification of brass instrument sounds using Convolutional Neural Network (CNN) combined with Mel-Frequency Cepstrum Coefficient (MFCC) feature extraction. This research aims to improve the accuracy of brass instrument sound recognition by utilizing CNN's ability to process audio data. Through experiments conducted with different audio durations and variations in CNN model architecture, this study evaluates the impact of dataset separation and model design on classification performance. The results show that dataset duration and CNN model architecture significantly affect classification accuracy, with the highest accuracy achieved in the scenario using 30 seconds of audio duration with an accuracy value of 84%. In addition, experiments varying the number of convolution layers in the CNN model show that the selection of the model architecture plays an important role in classification performance. Overall, this research contributes to advancing the field of audio classification by providing insight into the optimal dataset duration and model architecture for wind instrument speech recognition using CNNs.