• Align training data using the 800 syllable + baseline triphone system
• Pick 200 most frequent monosyllabic words from alignment
• Train syllables that have a minimum of 114 tokens
• Relabel alignments to reflect:
  • the 200 most frequent monosyllabic words
  • the 632 syllable models that have 114 training tokens
  • expand the 168 syllables (< 114 tokens) to word internal triphone baseform