In this paper, the attention is focussed on the importance of improving robustness against noise of all the pre-processing steps in a single-channel speech recognition system for operation in car environment. Different kind of parameter representation, noise reduction techniques and an automatic end-point detector were evaluated. Experiments were conducted by using a real car database and a Dynamic Time Warping (DTW) recogniser. At first, in an ideal condition (manual segmentation), the behaviour of two cepstral parameter sets and different speech enhancement techniques were compared. Afterwards, the performances of the recogniser using the best parameter were also evaluated in case of an automatic speech segmentation. Finally, an evaluation of the complete recognition chain with two different integrated Noise Reduction techniques was made.

Keywords: Signal Pre-Processing, Feature Extraction, Noise Reduction, End-Point Detection, Speech Recognition, Car Environment