Efficient Segmental Conditional Random Fields for Phone Recognition

Recently the initial attempt has been made to use segment-based direct models on their own for phone classification and recognition without the aid of an HMM lattice. This paper follows this line of research to further investigate these one-pass segmental direct models on phone recognition using posteriors as input. We make the first direct comparison between a frame-based system and a segmental system using the same base features, and explore the utilization of transition features in a direct segmental model for the first time. The results show that transition features can be very beneficial, particularly the ones surrounding the segment boundaries. In order to efficiently incorporate such features, we propose the Boundary-Factored SCRF, which reduces the time complexity of a Segmental Conditional Random Field (SCRF) to that of a frame-level CRF.