برچسب‌زني خودكار چندبعدي موسيقي ميكروتونال

automatic music tagging , Iranian dastgah music , music information retrieva‎l , deep learning , dynamic transition matrix , asymmetric cross‑attention mechanism

Multi-dimensional Auto-tagging of Microtonal Music

مهندسي كامپيوتر

The proliferation of online music streaming an‎d distribution platforms in the era of digital music has transformed the automatic analysis of musical content into a significant challenge at both industrial an‎d research levels. Functionalities such as fast search, personalized recommendation, an‎d playlist generation require the accurate identification of the semantic an‎d musical attributes of each track. Consequently, automatic music tagging has emerged as a central problem in music information retrieva‎l, playing a crucial role in the efficiency of these systems. Achieving these objectives necessitates the precise identification an‎d labeling of various characteristics of each musical piece, such as the modal system (e.g., dastgah in Iranian music) in which the piece is performed, the instruments employed, an‎d rhythmic properties. Given the vast volume an‎d diversity of musical content, as well as the specialized nature of music theory, manual annotation of these attributes is highly time-consuming, costly, an‎d practically infeasible. Therefore, there is a strong need for automated tools to perform this process. In response, the relatively recent research field of Music Information Retrieva‎l (MIR) has gained significant attention. However, the majority of existing studies in this domain have been developed based on Western music an‎d the equal-tempered 12-tone system. As a result, they face three fundamental challenges when applied to microtonal music, such as Iranian classical music: (a) the lack of suitable datasets, (b) the need for effective representations of melodic patterns, an‎d (c) the requirement for specialized architectures capable of efficiently processing such representations an‎d predicting labels. To address the first challenge, this dissertation introduces the Chakavak dataset, comprising 7,036 pieces of Iranian classical (dastgah-based) music annotated across three aspects—genre, instrumentation, an‎d textural features (with a total of 28 labels). This dataset is presented as the first benchmark dataset for Iranian classical music. To tackle the second an‎d third challenges, a novel transformer-based architecture, named Homayoun, is proposed. Leveraging innovative components such as a dynamic transition matrix an‎d an asymmetric cross-attention mechanism, the Homayoun architecture provides an effective representation of microtonal audio features guided by melodic patterns, enabling accurate label prediction for each musical piece. Experimental results demonstrate that the Homayoun architecture achieves Macro-F1 scores of 76.4, 76.2, 80.8, an‎d 80.7 on four microtonal datasets—Chakavak, Arabic genre, an‎d Turkish datasets (genre an‎d emotion), respectively—showing improvements of 14.5%, 5.8%, 4.9%, an‎d 4.3% over the best baseline models. Furthermore, in the context of Western music, eva‎luation on the GTZAN dataset yields a competitive accuracy of 77.4%, compared to the best baseline model (MERT) with an accuracy of 83.1%.

7

محمدهادي عيان بد

161359