Skip to content

Latest commit

 

History

History
37 lines (28 loc) · 2.02 KB

README.md

File metadata and controls

37 lines (28 loc) · 2.02 KB
Raw

Bass Line Labeling Framework

This is a demo project to evaluate various methods to transcribe and reproduce bass lines from audio recordings. In this project, we evaluated using various methods to transcribe and reproduce bass lines from audio recordings. The methods are: "baseline", "crepe_notes", "pesto_notes", and "basic_pitch". See the attached report for more details.

Installation

Ensure tensorflow is installed on your system. Then install the project using pip:

pip install -e .

Note that this repo only works with python==3.10

Download the example-files from here and place them in a directory called 'example-files' in the root of the project.

For tensorflow, since I'm on an M-series Mac, I had to install it using the following command:

python -m pip install tensorflow-macos

python -m pip install tensorflow-metal

Usage

    python scripts/label_and_generate.py

This will use the "crepe_notes" method to label the bass lines in the 'example-files' directory and generate the corresponding csv, midi, and audio files in the 'outputs/crepe_notes' directory. It uses the FluidSynth library to generate the audio files, so you'll need to have that installed on your system (follow the instructions here).

Evaluating with the Filobass Dataset

The report evaluated the methods using the "FiloBass" dataset. Download the dataset from here and place it in a directory called 'FiloBass' in the root of the project. You'll then need to convert all of the audio files to wav (I used FFMPEG). Then modify the label_and_generate script (uncomment the commented-out lines)

    python scripts/label_and_generate.py
    python scripts/evaluate.py

This will evaluate the performance of the "baseline", "crepe_notes", "pesto_notes", and "basic_pitch" methods on the FiloBass dataset. The results will be printed to the console.