A method for rescore aware pruning in speech recognition is disclosed. The method includes determining a set of possible words WP that are reachable in a CLG state from an active hypothesis HI. The CLG state may be mapped and dynamically assigned to CL state for determining WP and then remapped to the CLG. Given this information, a backtrace state associated with Hi in the CLG is examined and the set of words WR that will be rescored is determined. If WP and WR share an element, i.e., WP ⋂ WR ≠ ∅, then HI is retained and not pruned. Further, an additional beam may be used for increasing search space and include additional rescore-hypotheses.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.