Deep Griffin–Lim Iteration: Trainable Iterative Phase Reconstruction Using Neural Network

In this paper, we propose a phase reconstruction framework, named Deep Griffin-Lim Iteration (DeGLI). Phase is fundamental technique for improving the quality of sound obtained through some process in time-frequency domain. It has been shown that recent methods using deep neural networks (DNN) outperformed conventional iterative such as algorithm (GLA). However, computational cost DNN-based not adjustable at time inference, which may limit range applications. To address problem, combine structure GLA with DNN so becomes by changing number iterations proposed component. A training method independent inference also to minimize training. This method, sub-block denoising (SBTD), avoids recursive use and enables DeGLI single (corresponding one iteration). Furthermore, complex based on convolution layers gated mechanisms investigated its performance terms framework. Through several experiments, found significantly improved both objective subjective measures from incorporating DNN, was comparable those vocoders.