MULTIPLE COPIES OF PRONUNCIATION TREES

• A simple lexical tree is sufficient if no language model is used.
  
  
• For higher-order N-gram models, the linguistic state cannot be determined locally. History plays an important role. For example, for bigrams, a tree copy is required for each predecessor word.
  
  
• Efficient pruning can eliminate most of the unnecessary tree copies.
  
  
• To deal with tree copies, you can create redundant trees for each word context. However, this is expensive.
  
  
• Many of the active state hypotheses correspond to the same redundant unigram state because the language model probability cannot be applied until the next word has been observed.
  
  
• A successor tree approach can be used to eliminate this redundancy:
  
  
  
• However, for backoff language models, this isn't as efficient as it might seem (unless aggressive pruning is used to allow only a small subset of words to follow a given word).