Enhancing Automatically Discovered Multi-level Acoustic Patterns Considering Context Consistency With Applications in Spoken Term Detection.