Mid-level Harmonic Audio Features for Musical Style Classification

This repository contains the implementation of the musical style classification system proposed for the ISMIR 2022 submission "Mid-level Harmonic Audio Features for Musical Style Classification" (F. Almeida, G. Bernardes and C. Weiß).

Getting started

Install the required system dependencies:

• Python 3.9
• Pip package manager
• GNU Make
• Git LFS

Clone this repository and setup Git LFS. Please make sure the --recursive flag is passed so that all submodules are also cloned.

git clone git@github.com:fcfalmeida/style-ident.git --recursive
cd style-ident
git lfs install

Install project dependencies by running:

make install

Usage

Interaction with the system is done through the commands defined in the project's Makefile, which includes documentation on how to use them. For example, to compute TIV Basic features using the multiple resolution segmentation approach on the Cross-Era Piano dataset, run the following:

make res_feats dataset=crossera_piano pipeline=tis_basic_res

A pipeline represents the computation of a feature group using a given segmentation approach. Currently available pipelines are:

Harmonic Segmentation

• tis_complexity_segmented: TIV Complexity
• tis_basic_segmented: TIV Basic
• `harm_rhythm: Harmonic Rhythm

Multiple Resolutions Segmentation

• tis_complexity_res: TIV Complexity
• tis_basic_res: TIV Basic
• complexity: Tonal Complexity
• template_based: Template-based

Local Resolution Segmentation

• tis_complexity_local_res: TIV Complexity
• tis_basic_local_res: TIV Basic

Unprocessed datasets are placed under data/external and are provided in this repository for easier reproducibility of research results. The conducted study uses the following datasets:

• crossera_piano
• crossera_orchestra
• crossera_full
• crosscomp5
• crosscomp11