Exploring the Harmony of Deep Learning and Traditional Persian Music: A Journey of Modelling and Music Improvisation | by Behrad Gharedaghloo | Dec, 2023

In the vast expanse of music improvisation, where classical music and jazz genres particularly have been explored [1][2][3][4], my curiosity took me towards the enchanting world of traditional Persian music. Being an amateur musician with some experience in Persian music and its instruments, coupled with a background in engineering and science and in domains of machine learning and deep learning, I embarked on a hobby project that united my two passions: music and mathematics. The mission: to build a deep learning model capable of navigating the complex and nuanced realm of traditional Persian musical styles.

From the project’s outset, I recognized unprecedented challenges absent in prior studies. The abundance of microtones [5] in traditional Persian music, variations in modes, intricate motif structures, and substantial ornamentation — which is present in all traditional eastern music including Persian one — presented unique challenges. Additionally, differences in dataset size especially its digitized form underscored the distinctive nature of these challenges in the project. This necessitated the development of a dedicated tool capable of incorporating these nuances while constructing an inventive solution.

The genesis of this musical odyssey lay in the accurate collection and digitization of the “Radif” music [6], sourced carefully from the “Shour”[7] chapter of the Radif book by Dariush Talai [8]. This process, spanning over 10s of hours and extending across three months, served as a testimony to the dedication required to build the model. Digitizing the notes, then I defined dictionaries to encode and decode the notes, and later to transform these notes into one-hot vector sequences that would feed into the neural network model.

The implementation of the model was done using Keras in Python and by leveraging Long Short-Term Memory (LSTM) layers allowing the model to capture the intricate nuances of traditional Persian music. The overall code structure was obtained from Keras.io on character-level text generation [9] and the final model architecture was obtained from an article by Sigurður Skúli Published in Towards Data Science [10].

The journey faced its first challenge when the initial models failed to fit the input music sequences, revealing a significant underfitting issue. This setback necessitated a deep dive into the architecture of the model, and restructuring it that ultimately yielded a more robust model and more realistic music generations. The number of layers and the number of input notes before generating the next note were two important factors that significantly improved model accuracy. In its latest form, the possibility of generating every note correctly via feeding its prior notes was more than 95%.

In addition to the model calibration challenges, fine-tuning the temperature during the improvisation step became an art in itself — finding the delicate balance where the model neither looped into repetitive and predictable patterns nor veered into chaotic randomness.

The task of producing music files introduced a novel challenge. The existing libraries and MIDI files, crafted under the assumption of “Equal Temperament” that is the base for Western music. However, as mentioned above the Equal Temperament assumption is not valid in the face of the intricate scales of most eastern music forms including Arabic Maqam Music and traditional Persian music. To surmount this obstacle, I made my own sound library that would consider the microtonality by sampling note sequences using my Oud instrument. Using the library, the note sequences generated by the model (in improvising mode) were auralized into WAV files. Below are some examples of the generated music (and more are yet to come) on Soundcloud.