Emotional Melody Generation Model Based on Multi-scale Attention LSTM

Abstract

Aiming at the problems of ambiguous emotional representation and poor adaptability between melody details and overall structure in traditional melody generation models, this paper proposes a novel emotional melody generation model. Taking LSTM as the basic framework for extracting temporal features, a three-level multi-scale attention mechanism including fine-grained, medium-grained and coarse-grained attention is designed. It respectively captures detailed note features of melodies, local emotional features of phrases and overall structural emotional features of musical segments, which solves the defect that single-scale attention cannot balance the texture of melody details and global emotional consistency. An emotional dimension mapping mechanism is introduced to integrate the two-dimensional quantitative emotional parameters of valence and arousal into the feature encoding process, realizing the deep coupling of emotions and melody features. Comparative experiments are conducted on the public MAESTRO music dataset. The results show that the proposed model effectively addresses the issues of homogenized emotional expression and broken temporal dependence existing in traditional generation models.