Joint Learning of Emotion and Singing Style for Enhanced Music Style Understanding.

Area of Science:

• Music Information Retrieval
• Machine Learning
• Vocal Performance Analysis

Background:

• Music style understanding is crucial for applications like personalized recommendations.
• Existing research often treats emotion and singing style classification as separate tasks.
• The intrinsic relationship between vocal emotion and singing style is often overlooked.

Purpose of the Study:

• To develop a multi-task learning framework for joint emotion and singing style classification.
• To enable explicit knowledge sharing and mutual enhancement between the two tasks.
• To improve the robustness and accuracy of singing style analysis.

Main Methods:

• Implemented a multi-task learning framework to jointly model emotion and singing style classification.
• Evaluated the framework across diverse backbone architectures (Transformer, TextCNN, BERT).
• Conducted experiments on a self-constructed benchmark dataset using professional recording devices.

Main Results:

• Joint optimization consistently outperformed single-task learning approaches.
• The framework demonstrated stable performance improvements across different backbone architectures.
• Achieved state-of-the-art accuracy on both classification tasks within the benchmark dataset.
• Provided interpretable insights into the interplay between emotional expression and vocal style.

Conclusions:

• Jointly modeling singing style and emotion classification enhances analytical robustness.
• The proposed multi-task framework is generalizable and adaptable across various architectures.
• This approach offers a more comprehensive understanding of vocal performance characteristics.