A study on deep reinforcement learning based method for voice expression enhancement of opera singers

Abstract

Opera singer's voice expression is the soul of art. In reality, the singer's voice is susceptible to the interference of background noise, and the melody and rhythm of the voice are highly dependent on personal experience and state. In this paper, we propose a framework for enhancing the vocal performance of opera singers that integrates speech signal processing methods and deep reinforcement learning. The framework employs an improved convolutional transfer function generalized paraflap canceller algorithm (CTF-GSC) to suppress the non-coherent noise in the performance environment to achieve the noise reduction effect. A model based on two-layer unidirectional LSTM network and a model based on reinforcement learning AC algorithm are constructed to enhance the rhythm and melody of the generated music. The results show that the improved CTF-GSC algorithm effectively suppresses the correlated and non-correlated noise and enhances the vocal expression of the opera singer's voice. The generated music tunes are richer, and the probability of notes in the key reaches 98.8%. This study provides a new intelligent method for realizing the artistic performance of opera singing and opens up a new path for performance innovation in vocal art.