Luc LUSSIER and Alain BIEM
Channel Distortion Equalization for Robust Speech Recognition
Abstract:This report proposes an approach in speech recognition for handling mismatches in microphones that occurs between the training and the testing phase. Typically, it is assumed that the microphone used for training displays different characteristics from the microphone used in the testing phase. The proposed algorithm estimates a bias or distortion from the silence portion of the utterance. Assuming that a microphone acts as a filter on the incoming speech, the estimated distortion reflects the discrepancy between each microphones. Equalization is achieved by removing the estimated bias from the incoming speech.