Prototype-based MCE/GPD training for word spotting and connected word recognition

A straightforward application of PBMEC (prototype-based minimum error classifier) training to existing techniques for handling continuous speech is described. A novel MCE/GPD (minimum classification error/generalized probabilistic descent) loss function that can incorporate word spotting errors and other measures of symbolic distance between correct and incorrect categories is defined. Classification consists in a time-synchronous DTW (dynamic time warping) pass through a finite state machine; adaptation makes use of an A* based N-best algorithm and consists in propagating the derivative of the loss over the N best paths through the finite state machine. The key feature is that the loss function being optimized closely reflects the actual recognition performance of the system.<>