Implementation with PyTorch.

• Base model
  • LSTM using MFCC audio features
  • CNN(ref simplified version) with LPC features

• Python3
• PyTorch v0.3.0
• numpy
• librosa & audiolazy
• scipy
• etc.

• Scripts to run

  • main.py: change net name and set checkpoints folder to train different models
  • test_model.py: generate blendshape sequences given extracted audio features (need audio features as input)
  • synthesis.py: generate blendshape directly from input wav (need arguements of input audio path)

• Classes

  • models.py: Classes with LSTM and CNN (simplified NvidiaNet) model.
  • models_testae.py: Advanced models with audoencoder design.
  • dataset.py: Class for loading dataset.

• Input preprocessing

  • misc/audio_mfcc.py: extract mfcc features from input wav files
  • misc/audio_lpc.py: extract lpc features
  • misc/combine.py: combine certain audio feature/blendshape files to obtain a single file for data loading

To build your own dataset, you need to preprocess your wav/blendshape pairs with misc/audio_mfcc.py or misc/audio_lpc.py. Then combine those feature/blendshape files misc/combine.py to a single feature/blendshape file.

Modify main.py. Set model to the one you need and also specify checkpoint folder.

• Both test_model.py and synthesis.py can be used to generate blendshape sequences.
  • test_model.py accepts extrated audio features (MFCC/LPC).
  • synthesis.py takes raw wav file as input
  • State the arguments and it will produce a blenshape test file.