Reproducibility repository

With this repository we aim to make the ISMIR 2018 paper Music detection in broadcast audio recordings: a non-binary approach with relative loudness annotations as reproducible as possible.

Getting started:

1. Clone this repository

git clone https://github.com/BlaiMelendezCatalan/ISMIR2018.git

2. Download audio (and optionally Lidy's features):

3. Extract audio.zip, estimations.zip, annotations.zip and features.npz to the corresponding folder:

NOTE: estimations.zip and annotations.zip can be found in comparative_analysis/

Feature extraction and model training and testing

4. Follow the READMEs in the folder of each author (in comparative_analysis/algorithms/) to obtain their estimations for the testing dataset. The complete process including feature extraction, the training of the model and its testing can only be done for the Tsipas algorithm. For Lidy's and Marolt's algorithms only the testing part is available as the algorithms' code is not public. The result of the process for each author is already available in comparative_analysis/estimations/author/raw_estimations/.

Format estimations

5. Follow the README in comparative_analysis/format_estimations/ to format the estimations of each algorithm. The result of the formatting is already available in comparative_analysis/estimations/author/formatted_estimations/

Plot results

6. Follow the README in comparative_analysis/evaluation/ to obtain the plots in the paper.