Neural network development for noise reduction in robust speech recognition

Speech recognition systems with small and medium vocabulary are used as natural human interfaces in a variety of applications. To make such a system more robust, the development of a neural network based noise reduction module is described. Using standard feedforward networks, several topologies have been tested to learn about the properties of neural noise reduction. For the development of a sufficiently robust nonadaptive system, information about the characteristics of the noise and speech components of the input signal including context information was taken into account. The focus is on the stepwise experiment-oriented improvement of a basic linear neural noise reduction network. The isolated word recognition system and the database used for the experiments are described. Results from different noise reduction networks are given. To test their robustness, simulations with varying input signal characteristics were made and are discussed