• single-speaker means are shifted from model means (significantly for state 3)
  
  
• single-speaker variances are narrow compared to the model variances
  
  
• we want to adapt the model means and variances to match the single-speaker data