search

ljuvela / GlottDNN

GlottDNN vocoder and tools for training DNN excitation models
Apache License 2.0
32 stars 4 forks source link

TODO: DNN training block #12

Closed mairaksi closed 5 years ago

mairaksi commented 8 years ago

Theano data provider Python script Merlin compatibility?

mairaksi commented 8 years ago

Theano wrapper: -Should this include the parameter extraction of training data? E.g., user specifies the directory from which the vocoder parameters are extracted, and after the extraction, the script commences to train the DNN.

Inputs: -Wanted vocoder parameters -Desired DNN architecture -Folder containing parameter files (conventional + PAF) / wav files (if training data is to be extracted simultaneously) -Training parameters (e.g., max amount of training data, data normalization settings, number of epochs, learning rate, etc.)

Outputs: -Vocoder parameters (+PAF) if extracted -Binary format dnn-weights + min-max values in .dnnData file -Libconfig format .dnnInfo file (ASCII) containing the relevant information -Maybe an info dump about DNN training? (e.g., number of epochs, training + test errors, etc.)

ljuvela commented 5 years ago

Let's keep the DNN excitation model code to minimimum in this repo. I think the most convenient way is using external libraries for training and running the models.

Check these as a starting point https://github.com/ljuvela/ResGAN https://github.com/ljuvela/multiscale-GAN