• computational cost
  • both MMI and MCE are limited to small vocabularies if implemented in entirety
  • approximations include Viterbi alignment and N-best paradigm
  
  
• optimization techniques
  • choice of optimization - minimum, gradient descent, Newton's method etc.
  • step size for update of parameters needs to be empirically determined
  • parameter-wise step size often employed
  • choice of loss functions for MCE