This repository contains python files to compute pitch distributions aligned with the respected tonic frequencies of recordings and to perform supervised mode recognition. It uses culture-specific algorithms for Turkish Makam Music to obtain pitch distributions. For supervised mode recognition, a Multi-Layer Perceptron (MLP) model is used. The proposed method outperforms previous works on Turkish Makam Recognition Dataset.
Please cite the publication below, if you use the repository in your work:
F. Yesiler, B. Bozkurt and X. Serra. "Makam Recognition Using Extended Pitch Distribution Features and Multi-Layer Perceptrons". In 15th Sound and Music Computing Conference, Limassol, Cyprus, 2018.
This repository is mainly created to provide a straightforward access to pitch distribution computation and supervised mode recognition tasks. Currently, the algorithms for pitch extraction and tonic frequency estimation are culture-specific for Turkish Makam Music. An extension that includes other music traditions such as Hindustani Music and Carnatic Music will be the next step of this work.
Please refer to the PitchDistAMR-Demo.ipynb for a demonstration of this repository.
To compute pitch distributions, following cases can be considered:
1-Mode Information is Known
If the mode information is known, it should be included in "annotations.json" file located as:
pitchdistamr/data/your_data/annotations.json
"annotations.json" file includes:
- Mode of the recording
- Name of the recording
- Tonic Frequency of the recording (Optional)
An example annotations.json file:
[
{
"mode": "mode_1",
"name": "rec_1",
"tonic": 440.0
},
{
"mode": "mode_2",
"name": "rec_2",
"tonic": 220.0
}
]
When the mode information is known, the pitch distributions can be computed with:
- Extracting the pitch values of predominant melodies in the recordings or using the already extracted pitch values
- Estimating the tonic frequency of the recording or using the already estimated tonic frequencies included in the "annotations.json" file
The recordings or the pitch files should be organized regarding their modes. An example for the required file structure can be:
pitchdistamr/data/your_data/annotations.json
pitchdistamr/data/your_data/data/mode_1/rec_1.wav(rec_1.pitch, if the pitch files are to be used)
pitchdistamr/data/your_data/data/mode_2/rec_2.wav(rec_2.pitch, if the pitch files are to be used)
2-Mode Information is Unknown
If the mode information is unknown, the data directory should contain all the files. An example for the required file structure can be:
pitchdistamr/data/your_data/rec_3.wav(rec_3.pitch, if the pitch files are to be used)
pitchdistamr/data/your_data/rec_4.wav(rec_4.pitch, if the pitch files are to be used)
For supervised mode recognition, following cases can be consired:
1-Testing the method on recordings with known modes
After obtaining pitch distributions and modes as csv files, two use cases can be considered:
- Performing cross validation to tune the hyperparameters of the MLP model, using the obtained hyperparameters to create an MLP model, training the model with 90% of the data (training subset) and evaluating the trained model on 10% of the data (test subset).
- Creating an MLP model with specified parameters, training the model with 90% of the data (training subset) and evaluating the trained model on 10% of the data (test subset).
2-Identifying the modes of recordings using the pre-created model
After obtaining pitch distributions of the recordings as a csv file, the model trained with the entire dataset can be used to predict the modes of respective recordings.
This repository is written with Python 3. The list of libraries that are required are below:
numpy
scikit-learn
matplotlib
scipy
Essentia
The instructions for installing Essentia can be found here
An alternative way for using this repository is to install 'Docker'. Docker provides an easy access to a virtual environment with desired libraries. All the necessary libraries for this repository are installed and compiled in 'MIR-Toolbox', a docker image created for Music Information Retrieval applications.
The required installation steps are:
-
Install Docker Compose.
-
Clone this repository as follows:
git clone https://github.com/furkanyesiler/pitchdistamr.git
In a terminal/console window, change to this directory
On MacOS or Windows, run:
docker-compose up
On Linux, run the following (this command ensures that any files you create are owned by your own user):
JUPYTER_USER_ID=$(id -u) docker-compose up
Then accesss http://localhost:8888 with your browser and when asked for a password use mir
Furkan Yesiler furkan.yesiler@gmail.com
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
