基于图傅里叶变换的语音增强算法研究

Research on Speech Enhancement Algorithm Based on Graph Fourier Transform

摘要: 在语音增强过程中，人们往往采用语音信号的频谱信息作为特征输入，再进行进一步的训练增强处理。最为常见的便是对语音信号进行短时傅里叶变换后取其幅度频谱作为特征输入，在语音恢复阶段，则将含有噪声语音的相位信息作为增强语音的相位信息进行语音的重构。但是，这一做法必然导致相位信息的缺失。本文提出将图傅里叶变换(GFT)分别与非负矩阵分解(NMF)算法以及全卷积神经网络(FCNN)模型相结合来实现含有噪声语音的增强，实验表明，图傅里叶变换–非负矩阵分解算法在语音增强上与短时傅里叶变换–非负矩阵分解算法表现相当，基于图傅里叶变换–全卷积神经网络的语音增强相较于基于短时傅里叶变换–全卷积神经网络的语音增强有更为优异的性能。

Abstract: In the process of speech enhancement, people often use the spectral information of the speech signal as the feature input, and then carry out further training enhancement processing. The most used is to perform a short-term Fourier transform on the speech signal and take its amplitude spectrum as feature input, and in the speech recovery stage, the phase information of the noisy speech is used as the phase information of the enhanced speech for speech reconstruction. However, this practice inevitably leads to the absence of phase information. In this paper, it is proposed to combine the graph Fourier transform with the non-negative matrix factorization algorithm and the fully convolutional neural network model to realize the enhancement of noisy speech, and the experimental results show that the performance of graph Fourier transform-non-negative matrix factorization algorithm is comparable to that of the short-term Fourier transform-non-negative matrix factorization algorithm in speech enhancement, and the speech enhancement based on the graph Fourier transform-fully convolutional neural network has better performance than the speech enhancement based on the short-time Fourier transform-fully convolutional neural network.