Pengaruh Attention Pooling dan Metode Ekstraksi Fitur Terhadap Performa LSTM dalam Klasifikasi Emosi Musik Piano

Abstract

Music serves as a medium for emotional expression through acoustic elements such as pitch, rhythm, and timbre. This study focuses on music emotion classification based on solo piano instruments without vocal elements in the field of Music Emotion Recognition (MER). The objective of this research is to analyze the effect of pooling mechanisms and to compare the performance of MFCC and Log-Mel Spectrogram feature extraction methods on an LSTM model in classifying music emotions into four categories based on Thayer's quadrant model. The results demonstrate that the LSTM model with Attention Pooling consistently outperforms the LSTM model with Mean Pooling across all evaluation metrics, with accuracy improving from 63.97% to 66.91%, macro precision from 64.03% to 68.24%, macro recall from 64.53% to 66.65%, and macro F1-score from 64.14% to 66.81%. Furthermore, a comparison of feature extraction methods on the LSTM architecture with Attention Pooling reveals that the model based MFCC features outperforms the model based on Log-Mel Spectrogram features across all evaluation metrics, achieving a macro F1-score of 66.81% compared to 60.54%, with an accuracy gap of 5.15%.