AI MUSIC : Deep Network-Based Music Composition and Singing Voice Synthesis

Abstract

In this paper, we propose a novel deep network model for music composition and singing voice synthesis being studied in the artificial intelligence music field. In the first chapter, we propose a Multi- Objective Evolutionary Algorithm (MOEA) method that applies a deep network generation model to solve high-dimensional and complex problems such as musical composition. The proposed method demonstrates the scalability of the model and the validity of the proposed technique through two experiments on the knapsack problem and music generation. In the second chapter, we propose a non-autoregressive transformer based model suitable for Singing voice synthetics (SVS), which requires a long sequence. This paper proposes an improved SVS model compared to the existing Transformer model using LHS attention. The proposed method performs extensive experiments on sequences of various lengths, showing that it is memory efficient, has a faster inference speed, and generates good quality speech compared to conventional SVS models.