Automatic Speech Recognition in Adverse Acoustic Conditions

Automatic Speech Recognition (ASR) technology has reached maturity to the extent that it can be used successfully in various applications. However, it is by no means the “solved problem ” that some marketing campaigns are promoting it to be. One o f the biggest challenges that operational ASR systems are faced with, is to maintain recognition performance in adverse acoustic conditions. The training procedures o f most ASR systems yield recognisers with a relatively rigid image o f the world: Only those acous­ tic variations that actually occurred in the training data are accounted for. Since training data is usually clean (in the sense that care is taken to avoid noisy recording environments, channel noise, etc.), noise sources which are present when the system is operational result in a mismatch between the training and the test conditions. Such a mismatch may reduce recognition performance quite significantly. The aim of this research is to determine the extent to which the robustness o f ASR systems against mismatched training and test conditions may be increased using acoustic backing-off as an im­ plementation o f Missing Feature Theory.