Engineered Acoustic Features and Neural Networks for Music Emotion Recognition: an Empirical Comparison on the Deam Dataset
Year: 2026
Type: article
Abstract: This study compares engineered acoustic features with spectrogram based convolutional neural networks for static Valence and Arousal prediction on the 1,802-song DEAM dataset. Random Forest and XGBoost use 520 dimensional openSMILE features, while the CNN learns from 30-second log-mel spectrograms. The evaluation uses five saved song-level train, validation, and test partitions. The original XGBoost baseline achieves mean RMSE of 0.780 for Valence and 0.793 for Arousal, below every tested CNN configuration on both targets in all five runs. Increasing the CNN to 101,506 parameters lowers its mean RMSE to 0.837 and 0.855, respectively; a further increase to 403,714 parameters gives little additional benefit. A ten-configuration sweep for each tree model places the original settings among several competitive choices. Auxiliary analyses examine feature importance, prediction averaging, and linear recalibration. These results support openSMILE features with XGBoost as a practical baseline for this song-level prediction task.
Author: Lyu, Xiaoyu
Institution: California State University, San Bernardino
Language: English
Topic: Music and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225283216