GitHub's Polymath is a platform utilizing machine learning (ML) to convert any music library, whether sourced from Hard-Drive or YouTube, into a production-ready sample-library. The platform functions automatically by separating songs into stems such as bass, beats, etc., standardizing them to the same tempo and beat-grid, analyzing it's musical structure, and converting audio to MIDI. The goal is to simplify the workflow for music producers, DJs, ML audio developers, and ML developers who need to create a large music dataset for training generative models.
Polymath's search capability offers effortless access to a searchable sample library with related tracks, which enables users to easily combine different elements from various songs and create unique new compositions. The tool also allows for integrating stems from different tracks seamlessly into a Digital Audio Workstation in record time. Additionally, Polymath simplifies the process for analyzing a song's key, pitch, timbre, loudness, and information retrieval and processing, all possible through the platform's use of various neural networks.
Polymath's neural networks include DEMUCS, which facilitates Music Source Separation. SF_Segmenter is used for the Music Structure Segmentation/Labeling, while Crepe detects music pitch tracking, and Basic Pitch used for music to MIDI transcription. Lastly, Pyrubberband aids in music quantization and alignment.
Polymath is available via GitHub and requires the installation of the following software: ffmpeg, python version >=3.7, and <=3.10, Demucs Neural Net, sf_segmenter, Crepe Neural Network, Basic Pitch Neural Network, and LIBROSA. The platform is released under the MIT license.