• issues
  • which unit do we apply MCE to -- sentence, words or phones
  • zero-one loss function inappropriate for high WER
  • search space reduction needed -- N-best paradigm
• word/string level loss
  • define discriminant function for target class (a word in this case)
  • define the misclassification measure and loss as defined previously
  • use N-best paradigm and accumulate statistics for each class
  • MCE will push the models that are part of the misrecognized word away from the models that form part of the correct words
• computations
  • discriminant function related to the probability of the state sequence
  • loss related to discriminant function
  • using chain rule and GPD update parameters of the models