The proposed work introduces a new type of auto-tagging task, called instrument role classification. We discuss why the task is necessary, especially under the setting of making electronic music. Loop-based music and the general background regarding the creation of this specific style is also introduced. We approach this task with a previously introduced Convolutional Neural Network architecture, the HarmonicCNN (HCNN). A new use case of this method is presented by using the Freesound Loop Dataset (FSLD), emphasizing its learning efficiency under limited data. Finally, we present baselines to highlight these advantages.
To obtain the dataset used in this work, please refer to FSLD for the data.
And please cite this paper if you use this dataset:
@inproceedings{ramires2020,
author = "Antonio Ramires and Frederic Font and Dmitry Bogdanov and Jordan B. L. Smith and Yi-Hsuan Yang and Joann Ching and Bo-Yu Chen and Yueh-Kao Wu and Hsu Wei-Han and Xavier Serra",
title = "The Freesound Loop Dataset and Annotation Tool",
booktitle = "Proc. of the 21st International Society for Music Information Retrieval (ISMIR)",
year = "2020" }
To use the pretrained model of this work, set path to data and run loop_eval.py
- Run
data_preprocess/data.py
to get your audio data for training in npy format. - Run
data_preprocess/split.py
to split all data into train/valid/test. - Run
data_preprocess/check_length.py
to make sure the segments are all in the 3 seconds frame. - In
main.py
, make sure to change the data path to the directory of your npy files, and you can start training! - Lastly, run
eval.py
to check out the classification results.
[1] Joann Ching, Antonio Ramires, Yi-Hsuan Yang. "Instrument Role Classification: Auto-tagging for Loop Based Music" (KTH'20)