A review of robust techniques for the analysis of degraded speech

The severe effects of additive noise on automatic recognition systems has resulted in proposals of many different strategies. The paper reviews these strategies and loosely classifies them according to the stage at which the noise compensation occurs. the authors comment on the difficulty of making direct comparisons, given various inconsistencies, eg whether full knowledge of the noise degradation is known, either at the testing or the modelling stage. They suggest standards based on those criteria which define boundaries, and which would go some way to making comparisons more meaningful.<>