A Feed-forward Neural Network trained to learn a low-pass filter using TFCNNv3.
Originally I used Tensorflow Keras to train the original model in this repository. In that version the training sets are split into 9 sample chunks and a FNN is trained to take in 9 samples and output 9 filtered samples, obviously this is not great for continuity but it was easy to implement using only the high-level Python API for Tensorflow Keras, the original idea was to take 9 samples as input and output one sample and then shift the input one sample forward so that the input was always one sample increments with its four neighbours either side and that is what I have implemented in this version and TFCNNv3 made that job a lot easier for me.
The output is still not great and to be honest not that different from the original but the frequency spectra graph now looks a lot cleaner, in the original Keras version you could see the effect of the rigid 9 input sample chunks in the output frequency spectra.
Technically it would be better if I normalised the input and output data to -1 to 1 and then scaled back to 0 to 255 in the quantisation step, but honestly, I don't think it would make that much difference. Although, I did a test, trained for 1 hour and 15 minutes and it's actually better with normalised inputs; main_normalised_input_output.c.
I have provided one last solution main_float32_training_inputs.c which takes native float32 raw audio exports from Audacity, these come pre-normalised between -1 and 1, and then expects an unsigned 8-bit input song.raw to produce a transformed unsigned 8-bit output song_output.raw.
The three solutions are basically the same source code with minor modifications, this could have been presented in one source file with #ifdef tags but to avoid confusion I felt it better to seperate the source files for the three methods of presenting/normalising the training data.
To generate your input files, I used Audacity, Generate > Noise, save that as train_x.raw as Unsigned 8-bit PCM, then apply a low-pass filter to it and save that as train_y.raw in the same format as last time. Then find a song, load it into Audacity, Tracks > Mix > Mix Stereo Down to Mono, then export that as song.raw as Unsigned 8-bit PCM again. Once these three files are placed in the same project directory as compile.sh you should now be able to execute compile.sh which will compile and execute the training process. The training process will generate the first neural transformation of your song.raw and output it as song_output.raw. You may find it useful to apply a Normalize filter on the song_output.raw in Audacity as usually the output will have a small amplitude scale.
Just chmod 0700 or similar and execute compile.sh you will need GCC installed, you can find GCC in most (if not all) package managers.