Improving Keyword Spotting with a Tandem BLSTM-DBN Architecture

We propose a novel architecture for keyword spotting which is composed of a Dynamic Bayesian Network (DBN) and a bidirectional Long Short-Term Memory (BLSTM) recurrent neural net. The DBN uses a hidden garbage variable as well as the concept of switching parents to discriminate between keywords and arbitrary speech. Contextual information is incorporated by a BLSTM network, providing a discrete phoneme prediction feature for the DBN. Together with continuous acoustic features, the discrete BLSTM output is processed by the DBN which detects keywords. Due to the flexible design of our Tandem BLSTM-DBN recognizer, new keywords can be added to the vocabulary without having to re-train the model. Further, our concept does not require the training of an explicit garbage model. Experiments on the TIMIT corpus show that incorporating a BLSTM network into the DBN architecture can increase true positive rates by up to 10 %.