MOTIVATION FOR LATTICE GENERATION

Multiple-pass decoding is time and memory efficient for large scale speech recognition. Lattice generation generates a word graph with the N-best hypothesis, thus can efficiently constrain the search space for the second pass of decoding -- lattice rescoring.

Lattices provide nice framework for multi-pass decoding, and lattice is a convenient structure to store N-best hypothesis.

ISIP decoder currently operates in the lattice rescoring mode.