1

I have been executing an open-source Text-to-speech system Ossian. It uses feed forward DNNs for it's acoustic modeling. The error graph I've got after running the acoustic model looks like this: Plot of training vs. validation error of acoustic modeling Here are some relevant information:

  • Size of Data: 7 hours of speech data (4000 sentences)
  • Some hyper-parameters:
    • batch_size       : 128
    • training_epochs  : 15
    • L2_regularization: 0.003

Can anyone point me to the directions to improve this model? I'm assuming it is suffering from over-fitting problem? What should I do to avoid this? Increasing data? Or changing batch-size/epochs/regularization parameters? Thanks in advance.

Arif Ahmad
  • 111
  • 1

0 Answers0