• 41.8% WER at 60xRT on the dev-test from JHU WS'97: competitive with comparable commercial systems
  
  
• Front-End: Mel Cepstral (12) + Energy + Delta + Acceleration 60+ hrs training, cepstral mean normalization
  
  
• Models: 3-state left-to-right HMMs with dedicated silence and inter-word silence models, 40 phones, cross-word context-dependent triphones
  
  
• Training schedule: Baum-Welch training
  • Flat-start
    
  • Monophone Training
    
  • Triphone creation
    
  • State-tying
    
  • Mixture Generation
    
  
  
• Decoding: time-synchronous decoding
  • trigram, word-internal lattice generation
    
  • trigram, cross-word lattice rescoring