Combining State-level and DNN-based Acoustic Matches for Efficient Spoken Term Detection in NTCIR-12 SpokenQuery&Doc-2 Task.