Time synchronous decoding for long-span hidden trajectory model

摘要: A time-synchronous lattice-constrained search algorithm is developed and used to process a linguistic model of speech that has long-contextual-span capability. In the algorithm, hypotheses are represented as traces include an indication current frame, previous frames future frames. Each frame can associated unit such phone or units derived from phone. Additionally, pruning strategies be applied speed up search. Further, word-ending recombination methods computation. These effectively deal with exponentially increased space.