The performance of Long Short-Term Memory model variants in categorizing audio data related to traditional music is assessed in this research. The LSTM variants compared include LSTM, Bidirectional LSTM, Gated Recurrent Unit, Attention LSTM, Hierarchical LSTM, and Residual LSTM. The dataset carries the audio recordings of three traditional musical instruments: Gendang, Pui-Pui, and Kacaping. Mel Frequency Cepstral Coefficients are used to extract audio data. We model each variant of LSTM by training and testing the audio. Their performance was evaluated on Accuracy, Recall, and F1-score, where Residual LSTM outperformed with an average accuracy of 99% and an F1-score of 100%. To avoid overfitting, the k-fold cross-validation technique was used, where the results also showed the superiority of Residual LSTM with a stable accuracy of 99%. This study recommends the Residual LSTM model as the correct variant for performing traditional musical instrument sound recognition.