Developing a benchmark for emotional analysis of music

Area of Science:

• Computer Science
• Music Information Retrieval
• Affective Computing

Background:

• The field of music emotion recognition (MER) has seen rapid growth, with numerous new methods and audio features developed.
• Comparing MER algorithm performance is challenging due to diverse data representations and limited public datasets.
• This research addresses these limitations by creating a novel dataset and benchmark for MER.

Purpose of the Study:

• To establish a standardized benchmark for music emotion recognition (MER).
• To release the largest available dataset for dynamic MER, named DEAM (Database for Emotional Analysis in Music).
• To analyze and compare the performance of various MER algorithms and feature sets.

Main Methods:

• Creation of the DEAM dataset, featuring dynamic valence and arousal annotations for 1,802 songs at a 2Hz resolution.
• Organization of the 'Emotion in Music' task within the MediaEval Multimedia Evaluation Campaign (2013-2015).
• Analysis of benchmark results from 21 participating teams, focusing on winning algorithms and feature sets.

Main Results:

• The benchmark attracted 21 active teams, contributing to a comprehensive performance analysis.
• Recurrent neural network (RNN) based approaches, when combined with extensive feature sets, demonstrated superior performance in dynamic MER.
• The study provides insights into effective feature sets and algorithmic approaches for MER.

Conclusions:

• The DEAM dataset and benchmark provide a valuable resource for advancing music emotion recognition research.
• Dynamic MER performance is significantly improved by utilizing RNNs and comprehensive feature engineering.
• Standardized evaluation and larger datasets are crucial for reproducible and comparable MER research.