Why did the WER drop by 2.4% ?

• Changing the number of states in the Monosyllabic word and Syllable models
• Combining multiple pronunciations into one Monosyllabic word model
  • THE ( _dh_ah _dh_iy _dh_ax )
  • FOR ( _f_er _f_ow_r )
  • TO ( _t_ax _t_uw)
• Separation of different Monosyllabic words with the same baseform
  • _n_ow ( KNOW, NO )
  • _d_ey_r (THERE, THEY'RE, THEIR )
  • _t_ax ( TO ) _t_uw ( TOO, TWO )
• Requiring minimum of 100 tokens to update model parameters accounts for 0.1%