• Speaker Normalization
  • create a common feature space where the interspeaker variance is minimized
    
  • VTLN - dynamic frequency warping
    
  • trying to make all speakers appear equivalent so a common model set performs well on all speakers
    
  
  
• Speaker Adaptation
  • the recognition system is adapted to maximize performance for a single speaker
    
  • can use either batch or incremental adaptation
    
  • MLLR, speaker clustering