BASELINE SYLLABLE SYSTEM



  • WS96 syllabified lexicon used to get syllable definitions

  • 9023 syllables required to cover words in the training data

  • Syllable durations obtained from triphone based forced alignment

  • Number of states in syllable model chosen to account for 2 frames of input data per state

  • All syllables trained upto 8 gaussians/state



  • Testing - 800 most frequent syllables chosen, which cover 90% of the training data.

  • Monophones used to fill up the test lexicon


  • Performance - 55.1% WER