Dynamic optimization method for personalized vocal training program driven by reinforcement learning

Abstract

This paper proposes a personalized vocal training program dynamic optimization method based on reinforcement learning to address the problems of static solidification and poor adaptability of traditional vocal training programs, as well as the lack of dynamic optimization capabilities in existing intelligent vocal systems, making it difficult to achieve personalized teaching. Based on the theory of vocal music, audio signal processing technology, and deep reinforcement learning algorithms, an end-to-end cloud multi-source data acquisition architecture is built to preprocess and extract features from singing audio, behavioral, and physiological data. A high-dimensional continuous state space and a discrete continuous mixed action space are constructed. Design a multi-objective weighted reward function and temporal enhancement module with improved PPO algorithm as the core, to achieve real-time dynamic adjustment of training content, difficulty, duration, and rest strategy.