Description

- For these source separation implementation, that are open source, try a test run of the training the model. This may include:
- Re-training the model (open source repo may have instructions for doing so)
- Testing the existing models on test data
- Reading the source files and noting specific features or methods of the implementation

Rough steps to take
- Loading the dataset
- Train Model
- Evaluate model using performance metrics
- Compare to performance of existing solutions

Acceptance Criteria

• Notebook file created to test and play with the implementation code
• Peer review of code
• Documentation outlining comparison to existing solutions

After we test the current implementations of source separations, we should use the knowledge that we have to see what works and what doesn't work in order to come up with a new model architecture that we can try to build.

Description

This document will provide a basic overview of the current state of other audio separation techniques. Sections may include:

• Common datasets
• Common tools
• Common models
• Usual steps
• How machine learning is implemented
  Create a file under the subfolder docs/problem_overview

Acceptance Criteria

• File is created
• Peer review

Description

Interface for command line tool which is provided input audio file. Checks input and sends to preprocessing. Returns output

Acceptance Criteria

• Interface can be used to output dummy data given input.
• Interface accepts command line arguments without error.
• Code is in pep8 format and linted

Description

• Create a notebook file to test and organize data collection methods.
• Investigate datasets: musDB, spotify, etc
• visualize data in different formats ex: waveform, spectrogram

Acceptance Criteria

• Notebook file created in notebooks/ folder
• Peer review

Description

Need a Sepformer model trained on MUSDB dataset and the model file uploaded. Use the training steps in the sepformer_implementation notebook to get started. Model should be trained until validation loss stops improving.

Acceptance Criteria

• Model trained until validation loss stops improving.
• Models file uploaded in accessible location.
• Model training notebook added to repo with comments

Description

Create the initial Readme.md file. Give a background of the project and its goals. Some sections may include:

• About the Project
• Goal
• Installation
• Usage
• License
• Acknowledgements

Acceptance Criteria

• Peer reviewed
• Initial readme created.

Description

Initialize the repo for ability to install as a python package. Add folders/files that are not already included such as:

• preprocessing
• training
• inference
• test
• setup.py
• requirements.txt
• __init__.py

Acceptance Criteria

• Peer review
• Example files removed

Description

Update project readme with up to date information on it's usage and any other info. Add all dependencies of project to requirements.txt file.

Acceptance Criteria

• Project can be installed and used based on requirements.txt and information in readMe

Description

Several source separation libraries exist that provide a good framework to work with datasets and models. These can be used to speed up development and allow for easier hosting/access to training data.

Some of these include:

• nussl
• asteroid
• pyfasst | based on FAAST

Acceptance Criteria

• Notebook created for each of the libraries
• Summary of the available features
  • Datasets
  • Model architectures
  • Modularity
  • Documentation sources

Description

Modules which receive input from interface, convert to format acceptable to model, and run inference using model. The module should then convert output back to audio form.

Acceptance Criteria

• Inference runs successfully with multiple different input files
• Code is in pep8 format and linted

Description

Determine how to train Sepformer Model with Stereo Input, if not possible Mono input instead.

Acceptance Criteria

• Notebook uploaded which demonstrates method used with comments
• If modified Sepformer model used, upload code to repo in pep8 form and linted

Description

- For these source separation implementation, that are open source, try a test run of the training the model. This may include:
- Re-training the model (open source repo may have instructions for doing so)
- Testing the existing models on test data
- Reading the source files and noting specific features or methods of the implementation

Rough steps to take
- Loading the dataset
- Train Model
- Evaluate model using performance metrics
- Compare to performance of existing solutions

Acceptance Criteria

• Notebook file created to test and play with the implementation code
• Peer review of code
• Documentation outlining comparison to existing solutions