macemoth / audio-compression-msc

Code for the "CADENCE: Compressing Audio Data by Exploiting Cyclical Elements" master's thesis at the University of St. Gallen

audio-compression-msc

audio-compression-msc
├── compressors (Prototype compressors)
│   ├── ltoco.py (Lossless topological compressor)
│   ├── toco.py (Lossy topological compressor)
│   ├── utils
│   │   ├── ArithmeticBitCoder.py
│   └── veco.py (Vector quantisation compressor)
├── data
│   ├── adagio.wav
│   └── ...
├── notebooks
│   ├── fractal.ipynb
│   ├── nn.ipynb
│   ├── tensor.ipynb
│   └── vector_quantisation.ipynb
├── recompressor (MP3 recompressor)
│   ├── ArithmeticBinaryDecoder.py (Byte-level compressor)
│   ├── ArithmeticBinaryEncoder.py (Byte-level compressor)
│   ├── ArithmeticBitCoder.py (Bit-level compressor)
│   ├── Frame.py
│   ├── ...
│   ├── MP3Predictor.py (Probability model for ArithmeticBitCoder.py)
│   ├── NaiveArithmeticBitCoder.py (Probability model for ArithmeticBitCoder.py)
│   ├── Recompressor.py
│   ├── arithmetic_model.py (Probability model for ArithmeticBinary(En,De)coder.py)
│   ├── main.py (Entry point for recompressor)
│   └── ...
└── requirements.txt

Installation

1. Install dependencies using pip3 install -r requirements.txt
2. Only required for compressors: Copy desired audio samples from data/ to compressors/

Topological Data Compression

1. cd compressors
2. Run e.g. python3 toco.py c monoadagio.wav to compress lossily, or python3 ltoco.py c monoadagio.wav to compress lossless
3. Run python3 toco.py d monoadagio.tc to decompress

Running instructions are printed when running python3 toco.py or python3 ltoco.py

If input file is not mono, it will be converted, indicating too high compression ratios.

Tensor factorisation

1. Install TTHRESH and the python dependencies pip3 install tensorly notebook
2. Run jupyter notebook and open the tensor.ipynb notebook

For working through the complete tensor.ipynb notebook, TTHRESH (available at https://github.com/rballester/tthresh) is required.

Vector quantisation

1. cd compressors
2. Run e.g. python3 veco.py c 4 adagio.wav to compress
3. Run python3 veco.py d adagio.vc to decompress

To run notebook

1. Install pip3 install notebook
2. Run jupyter notebook and open the vector_quantisation.ipynb notebook

Fractal image compression

1. Run pip3 install scipy notebook
2. Run jupyter notebook and open the fractal.ipynb notebook

Neural networks

1. Run pip3 install notebook
2. Run jupyter notebook and open the nn.ipynb notebook

MP3 recompressor

1. cd compressors
2. Run e.g. python3 main.py trance.mp3 to recompress

By default, the optimised MP3 probability model is used. To choose the order-0 arithmetic coder or change other aspects of compression, modify the Frame.py file and follow the comments.

This module reuses the MP3 decoder code from https://github.com/tomershay100/mp3-steganography-lib.