This talk will introduce untwist, a new open source python library for audio source separation.
Audio source separation is a research topic in signal processing that has seen significant development during the last few years. Example applications include speech enhancement, music remixing and karaoke.
While many software libraries are available for audio analysis and music information retrieval, software for audio source separation is still scarce. Most research code is released in the form of Matlab scripts, while user-oriented software is implemented in C/C++.
untwist is a scientific python library developed in the context of audio source separation research, which can also be used for interactive prototypes. Many state-of-the-art algorithms, such as Non-negative Matrix Factorization (NMF), Robust Principal Component Analysis (RPCA), Harmonic-Percussive Source Separation (HPSS) and Deep Neural Networks (DNN) are implemented under a common object-oriented framework. Contrasting with existing available software, our library focuses on readability and ease of use, allowing the implementation of new pipelines with just a few lines of code. Development has now moved to github The talk will describe the main design challenges and proposed solutions for conducting audio signal processing research using scientific python, and will demonstrate some examples.