Synthesized-ComMU Dataset (SCM)

This is the repository of Synthesized-ComMU Dataset (SCM). This dataset was created by synthesizing midi from the ComMU[1] dataset with audio note files from the NSynth[2] dataset. Links to both datasets are provided below.

ComMU Dataset [Paper] [Dataset]
NSynyh Dataset [Paper] [Dataset]

MIDI to Audio Rendering

Rendering the midi I used into audio can be found [here]

Dataset Overview

Number of Samples: 130,440
- Train: 129,600
- Val: 240
- Test: 600
Features: Contains all metadata extracted from the midi provided by ComMU, as well as information about the NSynth presets used.
Classes: 6 perfectly balanced Track Role
- Main/Sub Melody, Accompaniment, Pad, Bass, Riff

Getting Started

Setup

Clone this repository
Download Datasets (ComMU, NSynth)
Install required packages
```
pip install -r requirements.txt
```

Preparation

The ComMU dataset can be preprocessed to achieve balanced Track role classes.
```
$ python preparation.py data-folder
```

After successful preprocessing, project tree would be like this,

  .
  ├── commu_meta.csv
  ├── commu_midi
  └── balanced
      ├── balanced_meta.csv
      ├── train
      │   ├── raw
      ├── valid
      │   ├── raw
      └── test
          └── raw

Augmentation

You can augment training data provided by the [ComMU-code]. The augmentation process will only involve the training data.
```
$ python preprocess.py --root_dir ./data-folder/balanced --csv_path ./data-folder/balanced/balanced_meta.csv
```

MIDI to Audio

```
$ python synthesize.py Nsynth-dir data-folder output-dir
```

Audio Output Results

Output audio file name is same as the MIDI file with the preset and source information added.
- 'midifile-name_preset_source.wav'
The files in the /output folder looks like this

output_dir/audio/train/midifile01_030_1.wav
output_dir/audio/train/midifile02_002_0.wav
output_dir/audio/val/midifile03_random_random.wav

Output csv: synthesized_results.csv

id	instrument	preset	source
midifile01	keyboard	030	1
midifile02	guitar	002	0
midifile03	guitar	random	random

spear011/SCM-Dataset