• Re-estimate the HYBRID system with 4 passes Baum-Welsh
• Seeded from the 800 syllable models having 8 mixtures per state
• Calculate median durations for 200 Monosyllabic words and 632 Syllable models
  • Based on the relabeled alignments from the 800 Syllable + triphone* system
• Monosyllabic word models seeded with the most frequent syllable pronunciation
  • 14 monosyllabic words have more than 1 pronunciation
• 100 tokens needed to update model parameters
  • Effects only triphone models
  • Monosyllabic words have at least 279 training tokens
  • Syllables have at least 114 training tokens

* The baseline word internal triphone system