Imposing Hierarchical Browsing Structures onto Spoken Documents

This paper studies the problem of imposing a known hierarchical structure onto an unstructured spoken document, aiming to help browse such archives. We formulate our solutions within a dynamic-programming-based alignment framework and use minimum errorrate training to combine a number of global and hierarchical constraints. This pragmatic approach is computationally efficient. Results show that it outperforms a baseline that ignores the hierarchical and global features and the improvement is consistent on transcripts with different WERs. Directly imposing such hierarchical structures onto raw speech without using transcripts yields competitive results.