Musical Emotion Recognition with Spectral Feature Extraction based on a Sinusoidal Model with Model-based and...

Area of Science:

• Signal Processing
• Computational Auditory Scene Analysis
• Affective Computing

Background:

• Audio signal analysis often relies on standard spectral features.
• Characterizing spectral shapes is crucial for understanding audio signal properties, especially timbre.
• Accurate emotion recognition from audio, particularly in music, is an ongoing challenge.

Purpose of the Study:

• To propose a novel method for extracting spectral features from audio signals based on a sinusoidal model.
• To evaluate the effectiveness of these novel features in predicting emotional dimensions (arousal and valence).
• To assess the contribution of these features in improving the accuracy of emotion recognition models.

Main Methods:

• Extraction of spectral features using a sinusoidal model, focusing on spectral peaks in frequency sub-bands.
• Utilizing prediction models such as principal component regression, partial least squares regression, and deep convolutional neural networks (CNNs).
• Evaluation of feature performance in predicting arousal and valence levels.

Main Results:

• The proposed spectral features capture additional spectral information beyond common baseline features.
• Experimental results demonstrate the utility of the novel features in characterizing spectral shapes.
• The inclusion of these features shows potential for reducing the prediction error rate for emotional dimensions.

Conclusions:

• The novel sinusoidal model-based spectral features offer valuable insights into audio signal characteristics.
• These features enhance the prediction of emotional dimensions, particularly valence, by better representing audio timbre.
• The proposed method contributes to improving the accuracy of music emotion recognition systems.