STACK DECODING (A^* SEARCH)

If some heuristics are available to guide the decoding, the search can be done in a depth-first fashion around the best path. We can avoid wasting computation on unpromising paths via time synchronous decoding. Such a heuristic function is very difficult to attain in speech recognition since it must combine elements of acoustic and language model scoring. Stack decoding is a variant of tree search.

• Note that the Viterbi search finds the optimal state sequence while stack decoding focuses on the optimal word sequence.
  
  
• The search process can be summarized as follows:
  
  
  • Add all possible one-word sequences to the OPEN list.
    
    
  • Remove the best path from the OPEN list; all paths from it are extended, evaluated, and placed back in the OPEN list (sorted).
    
    
  • Continue until a complete path that is guaranteed to be better than any path on the OPEN list has been found.
  
  
• Two key operations:
  
  
  • Finding an effective heuristic function for estimating the probability of the "future" part of the path.
    
    
  • Determining when to extend the search to the next word/phone.