Effect of Feature Smoothing for Robust Speech Recognition

Abstract

One class of feature enhancement techniques improve features' robustness by performing temporal filtering to smooth the feature trajectories. While smoothing can enhance the features' robustness by reducing the intra-class variation of the features, it also compromises the features' discriminative power by reducing their inter-class distance. In this paper, we investigate the effect of feature smoothing on speech recognition performance. To evaluate how different degrees of smoothing will affect the performance, the speech features are low-pass filtered with different cut-off frequencies and then used for model training and recognition. From the experimental results, we have two observations: 1) the noisy speech needs more aggressive feature smoothing; 2) the large vocabulary Aurora-4 task prefers less smoothing than the small vocabulary Aurora-2 task.