ابراهیم صحافی زاده

Music Information Retrieval (MIR) provides computational methods for music analysis; however, most existing approaches to Persian classical music rely on audio-based signal processing. This paper proposes a symbolic framework for Dastgah classification using Natural Language Processing (NLP) techniques applied to MIDI representations. Using the IRMA dataset of Mirza Abdollah’s Radif, MIDI note sequences are encoded symbolically by mapping pitch–duration pairs to Unicode characters and treated analogously to text. SentencePiece tokenization is employed to capture recurring melodic patterns, and Word2Vec is used to learn vector embeddings of the resulting musical tokens. Fixed-length representations of entire pieces are obtained via statistical pooling over token embeddings. These representations are evaluated using classical machine learning classifiers, including Support Vector Machines, Logistic Regression, k-Nearest Neighbors, and Random Forests. Experimental results show that mean-pooled representations combined with a linear SVM achieve the best performance, with an accuracy and F1-score of 0.81 across 13 Dastgahs and Avazes. To the best of our knowledge, this work presents the first application of NLP techniques to symbolic Dastgah classification, offering a complementary alternative to audio-based approaches that aligns closely with the theoretical structure of the tradition.