Syllable Level Recognition

[SlowTurtleGamer 6606]

Hello! I am trying to set up a syllable level recognition system by modifying the generic digits formula and changing the config parameters for mfcc.conf, train_mono and etc. The transcriptions in data/train/text are syllable level, and in lexicon its <syllable> <phon1> <phon2>. I just want to ask if there are any settings I should be looking at that can significantly affect the error rate considering the change from word level to syllable level transcriptions. Thank you!

For some extra details: Dataset contains about 2 hours of data, and initial word-level models resulted in <5% error. Any suggestions and guidance will be appreciated. :>