140901 - Shi Yin

From cslt Wiki
Jump to: navigation, search

Accomplished this week

  • Found the bug for noisy-training and re-trained the error model.
  • Decoded on the right model and trained two new models such as white/cafe/clean 15_1 & white/cafe/clean 15_0.1.
  • Decoded on the speech-rate model with different data and analyzed the result.
  • Tested dnn_vad on various database and analyzed the distributions about training data and evaluation data.

Next Week

  • Getting the results for new noisy models.
  • Checking the labels of database for dnn_vad.
  • Training the noisy model for dnn_vad.
  • Extracting speech data of different ROS, construct a new test set.
  • Try to remove silence when compute ROS.