Context modeling using RNN for keyword detection

The authors present some experiments that show the capabilities of using recurrent neural networks (RNNs) in conjunction with hidden Markov models (HMMs) in the context of keyword spotting (KWS): the automatic recognition of a small set of keywords as they occur in unconstrained speech and/or noise. KWS is usually considered as a word recognition problem on continuous natural speech with no syntax model. However, in this work, simple finite-state grammars are introduced to generate keywords in unconstrained speech. In the proposed scheme a RNN performs both the decoding search of a null-grammar HMM network and the secondary process to determine true keywords or false alarms.<>