ML-driven 1-channel denoising system

Semester project for Sound and Music Computing (AAU SMC 2018)

Link to report: https://www.overleaf.com/project/5c94e7edf54b1f21e4c1c066

Link to datasets: https://drive.google.com/drive/u/0/folders/1MmtiRQF-33Gkx1ZJZ7vSDCzy-vqjaUw6/

Proposal

Implementation of a single-channel denoising system in voice detection applications. The system would be based on deep learning techniques (e.g. autoencoders).

Initially, a baseline model shall be chosen, and an evaluation procedure established. Subsequently, improvements comprising techniques drawn from relevant literature will be implemented into the baseline, and evaluated accordingly. The project will be carried out in an iterative fashion.

Should satisfactory performances be achieved within a subset of the timeframe, the following further developments could be considered:

• Embedded implementation
• De-reverberation
• Multiple channels (e.g. with headset)

Instructions

Setup

• Setup remote environment:

  conda create --name <env_name> --file spec-file.txt
  source activate <env_name>

• Every time the spec-file.txt is modified, update environment:

  conda install --name <env_name> --file spec-file.txt

  Model definition

  Create new model as in models/ as model_<model_name>.py.

• Implement a custom class
• Constructor takes model parameters
• Expose get_model() method which returns a keras.Model object
• Expose get_lossfunc() method with return a loss function taking x_pred and x_true as arguments
• Define encoder and decoder as separate models
• Name explicitly all layers in autoencoder model
• See model_example.py for reference
• Cite the source (paper, repo, etc) on top

  Usage

• List available commands: python main.py --help

Structure

Code

• main.py: scripts entry point
• scripts/: scripts for training a model, viewing results, and using encoder and decoder
• libs/: code dependencies for scripts
• models/: model architecture implementations
• notebooks/: jupyter notebooks for experiments and tests
• tools/: miscellaneous software tools
• Pipfile, Pipfile.lock, environment.yml: list of dependencies, used for setting up pipenv (local) and conda (remote) environments

  Text

• notes/: minutes from group meetings
• literature/: relevant papers sorted by category
• ext/: unsorted, mixed stuff

Christie Laurent & Riccardo Miccini, 2019