Methods for noise reduction applied to speech input systems

Systems for noise reduction are used in speech-transmission and speech-input systems. Systems for noise reduction either can be a separate module of a speech-input system or are integrated into the speech-input system. The pros and cons of both solutions are discussed. As far as the noise reduction is concerned, several approaches are known which can be classified as noise-suppression systems based on Wiener filter theory and as noise-compensation systems using the least-mean square algorithm for adaptation. Whereas in the first case only one input for speech plus noise is used, the latter method requires one or more additional reference channels with correlated noise signals and without any speech component. A more recent development takes adaptive beamforming for noise suppression into account. Furthermore, systems based on neural networks are used for speech recognition in a noisy environment. It is shown that in realistic applications, e.g. a speech input system installed in a car, the recognition rate of the speech input system can be improved significantly if the speech-recognition system and the noise-reduction system are designed for the specific noise environment