19 Nov 1999
Sphinx Speech Group, CMU-SCS (rkm@cs.cmu.edu)
Sphinx-3 Search Algorithm
·initialize start state of <S> with path-score = 1;
·for each frame of input speech {
·        evaluate all active HMMs; find best path-score, pruning thresholds;
·        for each active HMM {
·                if above pruning threshold {
·                        activate HMM for next frame;
·                        transition to and activate successor HMM within word, if any
·                        if word-final HMM and above word-pruning threshold
·                                record word-exit in BP table;
·                }
·        }
·        transition from words exited into initial state of entire lexicon (using the
·                LM), and activate HMMs entered;
·}
·find final </S> BP table entry and back-trace through table to retrieve result;