/praudio

Audio preprocessing framework for Deep Learning audio applications

Primary LanguagePythonMIT LicenseMIT

README

praudio provides objects and a script for performing complex preprocessing operations on entire audio datasets with one command.

praudio is implemented having Deep Learning audio/music applications in mind.

Operations are carried out on CPU. Preprocessing can also be run on-the-fly, for example, while training a model.

The library uses librosa as an audio processing backend.

How do I install the library?

You can install praudio both with pip via PyPi, and by cloning the praudio repo from GitHub.

For both approaches, it's advisable to use a dedicated Python virtual environment.

Installing from PyPi

Installing from PyPi is the easiest option. In the terminal type:

$ pip install praudio

Installing from GitHub

First, you should clone the repository from GitHub:

$ git clone git@github.com:musikalkemist/praudio.git

Then, move to the project root and, to install the package, type in the terminal:

$ pip install .

You can also use a rule in the available Makefile (see below):

$ make install 

To install the package in development mode use:

$ pip install -e .[testing]

You can also use a rule in Makefile:

$ make install_dev 

This will install all the packages necessary to run the tests, lint, type checker. It will also install the package in 'editable' mode, which is ideal for development.

Python version

praudio works in Python 3.6, 3.7, 3.8.

How do I preprocess an audio dataset?

The core of the library is the preprocess entry point. This script works with a config file. You set the type of preprocessing you want to apply in a yaml file, and then run the script. Your dataset will be entirely preprocessed and the results recursively stored in a directory of your choice that can potentially be created from scratch.

To run the entry point, ensure the library is installed and then type:

$ preprocess /path/to/config.yml

In the config.yml, you should provide the following parameters:

  • dataset_dir: Path to the directory where your audio dataset is stored
  • save_dir: Path where to save the preprocessed audio.
  • Under file_preprocessor, you should provide settings for loader and transforms_chain.
  • loader: Provide settings for the loader.
  • transforms_chain: Parameters for each transform in the sequence. of transforms which are applied to your data (i.e., TransformChain).

These config parameters are used to dinamically initialise the relative objects in the library. To learn what parameters are available at each level in the config file, please refer to the docstrings in the relative objects.

Check out test/config.sampleconfig.yml to see an example of a valid config file.

Package structure

The package is divided into a number of subpackages:

  • config
  • creation
  • io
  • preprocessors
  • transforms

config has facilities to load, save, and validate configuration files, which are used to specify the types of preprocessing pipelines to use.

creation has classes that are responsible to instantiate key objects in the library.

io contains facilities to load / save audio signals from / to files.

preprocessors features objects which are responsible to preprocess single audio files, from loading to storing, as well as, batch of files.

transforms contains a series of objects which manipulate audio signals, such as short-time Fourier transform, log, scaling.

What's the Makefile for?

The Makefile has a series of rules that can be used to ensure quality of the code, and automate repetitive tasks.

Linter

The project uses pylint. The linter helps enforcing a coding standard, sniffs for code smells and offers simple refactoring suggestions.

To run the linter type:

$ make lint

Typehint

The project uses mypy. mypy is an optional static type checker for Python. You can add type hints (PEP 484) to your Python programs, and use mypy to type check them statically.

To run the type checker type:

$ make typehint

Testing

The project uses pytest for unittests. Tests can be run in one go using coverage. This package suggests the percentage of code that is covered in unittests.

To run all the unittests type:

$ make test

Checklist

Checklist is a utility rule that runs the linter, type checker, and the test suite in one go:

$ make checklist

Clean

Use the clean rule to get rid of pyc files and __pychache__:

$ make clean

Dependencies

praudio has the following dependencies:

  • librosa==0.8.1
  • pyyaml==5.4.1
  • types-PyYAML==5.4.6

librosa is extensively used to extract audio features in transform objects.

Current limitations

The praudio preprocessors are capable of operating only on mono signals. This is a significant limitation if you are working in generative music. If you are using the library for audio / music analysis, this shouldn't be a problem.

Future improvements

  • Add audio augmentation / padding / cropping transforms.
  • Enable preprocessing of signals with multiple channels.
  • Turn transform parameters into full-fledged objects (e.g., STFTParams)
  • Instead of using a dictionary for configurations, instantiate parameter objects with validation
  • Implement different types of Savers / Loaders with factories to produce them.