140803 - Mengyuan Zhao

From cslt Wiki
Jump to: navigation, search
  • about sinovoice
  1. extract car-noise from records and start noise training
  2. further test delta-feature, it performs better, but stream model is worse than non-stream model. using delta-feature training non-stream model, now
  • dropout nnet experiment
  1. finish reading the paper. now, modifying kaldi code.