Sparse imputation for noise robust speech recognition using soft masks

In previous work we introduced a new missing data imputation method for ASR, dubbed sparse imputation. We showed that the method is capable of maintaining good recognition accuracies even at very low SNRs provided the number of mask estimation errors is sufficiently low. Especially at low SNRs, however, mask estimation is difficult and errors are unavoidable. In this paper, we try to reduce the impact of mask estimation errors by making soft decisions, i.e., estimating the probability that a feature is reliable. Using an isolated digit recognition task (using the AURORA-2 database), we demonstrate that using soft masks in our sparse imputation approach yields a substantial increase in recognition accuracy, most notably at low SNRs.