2021 Volume 25 Issue 6 Pages 233-237
We propose a method to enhance noise reduction performance by separating a speech spectrum into spectral envelopes and fine structures using cepstrum analysis and linear predictive coding (LPC) analysis, and removing noise using an autoencoder (AE). A technique for removing noise from the spectrum of noise-containing speech is to use AE to reconstruct the spectrum of speech through the latent variables of the speech. We focused on spectral envelopes and fine structures that constitute speech, and improved the independence between latent variables in AE to reconstruct the speech spectrum by separating them in advance. In this way, we confirmed that the performance of noise reduction was improved in exchange for a slight decrease in the reproducibility of speech spectra when cepstrum analysis was used. It was also confirmed that cepstrum analysis was superior to LPC analysis in noise reduction.
