Connectionist Subspace Decomposition for Speech Enhancement

In this chapter, a two-stage noise removal algorithm that deals with additive noise is proposed. In the first stage, a feedforward neural network (NN) with a backpropagation training algorithm is applied to match the uncorrupted information. In the second stage, the Karhunen-Loeve Transform (KLT) based subspace filtering is used to compensate for the destruction caused by the noise. This combination is motivated by the fact that neural networks have the ability to learn from examples, even from complex relationships (non-linear) between inputs and outputs, and that subspace filtering has demonstrated its effectiveness to perform noise reduction through an optimal representation of features.