MORTY: A Toolbox for Mode Recognition and Tonic Identification
|Title||MORTY: A Toolbox for Mode Recognition and Tonic Identification|
|Publication Type||Conference Paper|
|Year of Publication||2016|
|Conference Name||3rd International Digital Libraries for Musicology Workshop (DLfM 2016)|
|Authors||Karakurt, A., Şentürk S., & Serra X.|
|Conference Start Date||12/08/2016|
|Conference Location||New York, USA|
|Keywords||carnatic music, hindustani music, k-nearest neighbors classification, Mode recognition, Open Source Software, Ottoman-Turkish makam music, Pitch Class Distribution, Reproducibility, Tonic identification, Toolbox|
|Abstract||In the general sense, mode defines the melodic framework and tonic acts as the reference tuning pitch for the melody in the performances of many music cultures. The mode and tonic information of the audio recordings is essential for many music information retrieval tasks such as automatic transcription, tuning analysis and music similarity and automatic description of digital music libraries applied to these cultures. In this paper we present MORTY, an open source toolbox for mode recognition and tonic identification. The toolbox implements a generalized variant of two state-of-the-art methods based on pitch distribution analysis. The algorithms are designed in a generic manner such that they can be easily optimized according to the culture-specific aspects of the studied music tradition. We test the generalized methodology systematically on the largest mode recognition dataset curated for Ottoman-Turkish makam music so far, which is composed of 1000 recordings in 50 modes. We obtained 95.8%, 71.8% and 63.6% accuracy in tonic identification, mode recognition and joint estimation of the mode and tonic tasks, respectively. We additionally present recent experiments on Carnatic and Hindustani music in comparison with several methodologies recently proposed for raga/raag recognition. We believe that our toolbox would be used as a benchmark for future methodologies proposed for mode recognition and tonic identification, especially for music traditions in which these computational tasks have not been addressed yet.|
To test the generalized methodology we have gathered a dataset in github (link) composed of 1000 recordings in 50 modes. It is the largest mode recognition dataset curated for Ottoman-Turkish makam music so far.
Experiments on Ottoman-Turkish makam music
We test the generalized methodology systematically on the dataset described above. We obtained 95.9%, 71.4% and 63.2% accuracy in tonic identification, mode recognition and joint estimation tasks, respectively. The complete experiments are released in github (link).
The files storing the features, the training models, the test results and the evaluation exceed 1 GB, which github does not host due to file size constraints. These files are stored in Zenodo (link) instead.
Additional experiments on Hindustani and Carnatic music
Gulati, S., Serrà J., Ganguli K. K., Şentürk S., & Serra X. (2016). Time-Delayed Melody Surfaces for Rāga Recognition. 17th International Society for Music Information Retrieval Conference (ISMIR 2016).
Gulati, S., Serrà J., Ishwar V., Şentürk S., & Serra X. (2016). Phrase-based Rāga Recognition Using Vector Space Modeling. 41st IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2016). 66-70.