Constructing Acoustic Distances Between Subwords and States Obtained from a Deep Neural Network for Spoken Term Detection.