Start- and end-node segmental-HMM pruning
An efficient decoding algorithm for segmental HMMs (SHMMs) is proposed with multi-stage pruning. The generation by SHMMs of a feature trajectory for each state expands the search space and the computational cost of decoding. It is reduced in three ways: pre-cost partitioning, start-node (SN) beam pruning, and conventional end-node (EN) beam pruning. Experiments show that partitioning cuts computation by 20–25% for supervised training, and 40–50% for phone classification, without degradation in recognition accuracy; SN and EN beam pruning together give 80% reduction for embedded recognition on triphone SHMMs, with less than 0.1% degradation.
