• on average a 35% relative decrease in WER for supervised adaptation and 15% relative decrease for unsupervised
  
• similar results hold for larger data sets (e.g. Broadcast News, SWITCHBOARD)