DNAE-GAN: Noise-free acoustic signal generator by integrating autoencoder and generative adversarial network

Abstract

Linear predictive coding is an extremely effective voice generation method that operates through simple process. However, linear predictive coding–generated voices have limited variations and exhibit excessive noise. To resolve these problems, this article proposes an artificial intelligence model that combines a denoise autoencoder with generative adversarial networks. This model generates voices with similar semantics through the random input from the latent space of generator. The experimental results indicate that voices generated exclusively by generative adversarial networks exhibit excessive noise. To solve this problem, a denoise autoencoder was connected to the generator for denoising. The experimental results prove the feasibility of the proposed voice generation method. In the future, this method can be applied in robots and voice generation applications to increase the humanistic language expression ability of robots and enable robots to demonstrate more humanistic and natural speaking performance.