Signal subspace speech enhancement

Abstract

Speech enhancement intends to improve the quality of speech by using various algorithms. Quality states clarity, intelligibility, pleasantness and compatibility. In most of the speech enhancement processes, speech quality improvement is achieved with the suppression of background noise or by estimating the background noise. In this project we are going for suppression of background noise. First, decompose the noisy speech signal into noise subspace and signal-plus-noise subspace using Karhunen-Loeve transform (KLT). Then remove noise subspace and linearly estimate the clean signal from signal plus noise subspace for enhancing the speech signal. The noise is assumed to be additive and uncorrelated (i.e., white noise) with clean signal. Estimation of clean signal is performed on frame-by-frame basis using two perceptually meaningful criteria. First, the time domain constraint that minimizes the signal distortion while the energy of the residual noise falls below some threshold (i.e., wiener filter). Second, for a fixed spectrum of the residual noise the signal distortion is minimized. Signal subspace approach is proven to be optimal for large samples in linear minimum mean square error

sense, where the signal and noise are assumed to be stationary.