A speech enhancement method based on graph fourier transform domain diffusion generative model

Abstract

The application discloses a speech enhancement method based on a diffusion model generated by graph Fourier transform domain, utilizes a graph signal processing theory to construct an adjacency matrix for a speech frame, and obtains a graph Fourier transform basis through singular value decomposition; the STFT in SGMSE+ is replaced by the graph Fourier transform basis, a one-dimensional convolution module is used to extract GFT domain features of noisy speech; subsequently, a forward stochastic differential equation is defined in the GFT domain to introduce noise, and a deep neural network is used for fractional matching to solve the inverse SDE, and the GFT features of clean speech are gradually recovered; finally, one-dimensional transpose convolution modules are used for inverse graph Fourier transform, and enhanced time-domain speech is reconstructed. The application breaks the limitation of traditional time-frequency analysis, improves the representation ability of the model for the structure of speech, and further improves the quality and naturalness of the generative speech enhancement without significantly increasing the computational complexity.