FACTORED LANGUAGE MODEL PROBABILITIES

• Premise: We can improve accuracy and minimize resources if we can apply the language model score as soon as possible.
  
  
• Factoring LM probabilities across a tree is one such idea:
  
  
  
• We can embed these probabilities in the pronunciation tree:
  
  
  
• In practice, such ideas require strong emphasis on the language model to be effective (such as WSJ or NAB). Applications such as conversational speech, in which acoustic ambiguity is extreme, and acoustic scores tend to dominate decoding, do not benefit as much from such approaches.