jakobabesser / bassunet

BassUNet - U-Net based convolutional neural network for (jazz) bass transcription proposed by Abeßer & Müller in 2021

BassUNet

Reference

Algorithm for bass transcription (joint frame-level pitch and voicing estimation) using U-Net Fully Convolutional Networks

• J. Abeßer & M. Müller: BassUNet: Jazz Bass Transcription using a U-Net Architecture, Electronics, 10(6), 2021
  • https://www.mdpi.com/2079-9292/10/6/670

Run

We recommend you to install miniconda (https://conda.io/miniconda.html). You can create a suitable environment using

conda create --name bassunet python="3.6"
conda activate bassunet
pip install librosa tensorflow==1.15 "h5py<3.0.0"

You can run the bass transcription algorithm as shown in the bassunet.py file:

bun = BassUNet()
t, f0, onset, duration, pitch = bun.run(wav_file_name)

Example

Now you can run the transcription algorithm on a test file by calling

python bassunet.py

After running the transcriber on the test file ArtPepper_Anthropology_Excerpt.wav, the frame-level pitch estimates as well as the estimated note events are stored in two CSV files. Both can be imported into Sonic Visualiser as time-instance and note layers, respectively.

You can open the example SV project ArtPepper_Anthropology_Excerpt.sv for an example.

Enjoy.