• CORE 12 Project.
  Hardware speech sensor and its optimization based on evolutional methods
• Meta-parameter settings of LSTM language model optimized by evolutionary computation.
  Optimized meta-parameters
• CSJ recipe for Kaldi
  Runs large vocabulary continuous speech recognition using Corpus of Spontaneous Japanese (CSJ)
  The system is tuned by an evolutional algorithm using TSUBAME super computer.
  The averaged word accuracy for the CSJ test sets is more than 91%.
  • Avilable in the official Kaldi package under egs/csj
  • Tutorial for the Kaldi CSJ recipe (2016/2, 2016/9)
• Eye motion input based speech synthesis system for communication assistance
  Demo video MP4(5MB)
• husky2
  A speech recognizer purely based on functional programming
• CV/Ag-EM
  demo script
• Unsupervised CV adaptation
  demo script
• HTK based script for triphone modeling
  sample script

---

[back to home]