An investigation on DNN-derived bottleneck features for GMM-HMM based robust speech recognition

In recent years, deep neural network(DNN) has achieved great success when used as acoustic model in speech recognition. An important application of DNN is to derive bottleneck feature. In this paper, firstly we investigate the robustness of bottleneck features generated by three types of DNN structures on the Aurora 4 task without any explicit noise compensation. Secondly, we propose the node-pruning method to reconstruct DNN which generates a new type of deep bottleneck feature. Our experiments show that bottleneck features generated from the node-pruned DNN achieve a promising reduction in word error rate(WER) when compared to the other bottleneck features produced by the conventional types of DNN structures. In addition, the new approach using the node-pruned DNN structure can automatically obtain the compact layer which generates the bottleneck feature.