In this paper we present evidence to suggest that a wavelet based noise reduction technique, having been used predominantly in image proccessing, can be applied successfully to speech signals in order to improve speech recognition accuracy. The algorithm requires no retraining of acoustic models and is directly applied to the time waveform, thus it can be used as a preproccessing step for any recognition system. We test the technique on both additive and convolutional noise using two different speech recognizers.