Keras and PyTorch implementations of Low Delay filter banks, Modified Discrete Cosine Transform filter banks, and Pseudo Quadrature Mirror filter banks
The advantages of the neural network implementation are:
- The possibility to use the GPU in this way and speed up the computation,
- hence it might not need a specialized fast implementation,
- It can be easily integrated in AI-based programs, for instance for source separation, or as a starting point for further optimization or training.
It needs the Keras, Theano and PyTorch libraries, install it in Python2 with: sudo pip install keras sudo pip install Theano sudo pip install torch
in Python3: sudo pip3 install keras sudo pip3 install Theano sudo pip3 install torch
For details and examples of filter bank design, look at: https://github.com/TUIlmenauAMS/Jupyter_notebooks_AMS/tree/master/Audio_Coding
and:
https://www.tu-ilmenau.de/mt/lehrveranstaltungen/lehre-fuer-master-mt/audio-coding/
Gerald Schuller, May 2020.