Searching for a song is a necessity, where the copyright of the song is a significant concern. This study proposes a method to classify and identify songs based on specific features that the model learns from music data. Python and CNN programming languages are used to build the model. In the first process, support libraries are used to extract audio data from the computer in WAV format. The dataset A includes 100 songs without lyrics, while the dataset B includes 100 audio files with the same song name but played in different types of musical instruments. We randomly cut the original audio files into clips less than 10 s long because users often use a specific code to find the entire track. The original audio files are split into clips of different lengths in the training set, including 1, 3, 5, 10, 20, 30, 60, and 90 s. Next, the Short-Time Fourier Transform was used to convert the audio data to the frequency domain. Finally, a shallow Convolutional Neural Network (CNN) and a Fully Connected layer (FC) were used to perform song classification tasks. We found that data augmentation by dividing the entire song into small pieces based on length significantly improved classification performance compared to those not using this technique. This research positively contributes to the advancement of e-commerce music systems, where listeners can enjoy music conveniently and memorably.