Probabilistic compensation of unreliable feature components for robust speech recognition

Missing feature theory is well studied in robust ASR context, many works have been done on additive noise of different colors. These are based mainly on classical spectral subtraction and marginal density techniques. This paper addresses the problem of temporal distortion of feature components, that is all about time domain instead of frequency one. No specific noise model and extract computation needed. We showed that the digit words recognition rate is above 95%, given test samples are clean with 10dB white noise added to middle 30% portion of speech along the time axis.