Improved Bottleneck Feature using Hierarchical Deep Belief Networks for Keyword Spotting in Continues Speech

Bottleneck (BN) feature has attracted considerable attentions by its capacity of improving the accuracies in speech recognition tasks. Recently, researchers have proposed some modified approaches for extracting more effective BN feature, but these approaches still need further improvement. In this paper, motivated by both deep belief networks (DBN) and hierarchical Multilayer Perceptron (MLP), we propose hierarchical DBNs based BN feature and employed it for keyword spotting task. The hierarchical DBNs based BN feature is constructed with two DBNs in series which are sequentially trained. The first DBN outputs the posterior probabilities features, as well as the second DBN transforms the posterior probability features into a low dimensional representation with the information pertinent to classification through the BN layer. Experiments on hierarchical DBNs based BN feature is conducted with TIMIT dataset and using Point Process Model as the baseline system. Experimental results show that the hierarchical DBNs based BN feature is more robust and can achieve better accuracies than other features.