Phrase language models for detection and verification-based speech understanding

We propose a phrase language model that has two key features. First, the model is oriented for robust understanding of unconstrained speech. Second, it does not need a large task-specific training corpus. The basic idea is that we focus on the stable and significant patterns of variable-length phrase expressions rather than uniformly modeling the whole utterances, and then classify them into task-dependent portions and task-independent ones. While the task-dependent key-phrases are trained with a small amount of task-specific data, the task-independent model is constructed with other large corpora that are not necessarily related to the current task. The task-independent model extracts expressions specific to the dialogue style rather than the task domain, and complements the task-dependent key-phrase model to enhance the detection and verification performance.