• discriminative techniques
  • ML is not the most appropriate for HMM estimation
  • need to incorporate negative examples into optimization
• maximum mutual information
  • motivated from the definition of mutual information and conditional entropy
  • incorporates negative examples via a likelihood measure
  • difficult to go beyond applications with constrained grammars
  • N-best paradigm is an alternative
• minimum classification error
  • motivated by the theory of GPD
  • optimization via a set of loss functions
  • loss defined in terms of expected misclassification
  • can be applied to string and phone level
  • computationally expensive for large vocabulary applications