Feature enhancement error compensation for noise robust speech recognition

This paper presents an approach to feature enhancement error compensation for noise robust speech recognition. The conventional feature enhancement techniques estimate the enhanced clean speech from the noise corrupted speech for improving speech recognition performance under noisy environments. During speech feature enhancement process, undesired residual error is generated because of incomplete property of the noise reduction. We apply the switching linear dynamic transducer (SLDT) to compensate this residual error. The SLDT describes the sequence-to-sequence mapping in a systematic way and has been applied to stereo data based speech feature mapping for channel distorted speech recognition. We assume that feature enhancement is a channel. The proposed method shows recognition error reduction in Aurora 2 digit task and Aurora 4 large vocabulary task with the interacting multiple model.