Soft decision strategy and adaptive compensation for robust speech recognition against impulsive noise

This paper presents research on robust automatic speech recognition (ASR) in the presence of impulsive noise, which is usually caused by transmission errors or packet loss in network-based delivery of speech signals. A soft decision strategy is proposed by analyzing the degraded observation probabilities caused by impulsive noise. Based on the soft decision results, two compensation methods are developed. The first aims at suppressing the unreliable likelihood scores by flooring the observation probabilities (FOP) on sensitive feature components with an adaptive threshold. The second focuses on the recovery of corrupted features and the unreliability of reconstructed data can be further compensated by the flooring method. Evaluation results on the Aurora connected digits database show that the proposed methods significantly improve the recognition robustness against impulsive noise. For example at the occurrence rate of 50% in simulated impulsive noise environment, the accuracy is increased from 42.74% of the baseline to 85.35%.