PILOT EXPERIMENT: DATA SPLITTING IN STATE SPACE

baseline system: 800 syllables (8 mixtures per state) + triphones

extraction of trajectories:

clustering of training data for top 485 syllable models based on distance to original model: state acoustic likelihood
(Korkmazskiy et al. 1997)

splitting of distance data into two clusters along mean of distribution

new model topologies:

two separate paths, 4 mixtures per state, different numbers of states