Abstract
Automatic raga recognition is one of the fundamental computational tasks in Indian art music. Motivated by the way seasoned listeners identify ragas, we propose a raga recognition approach based on melodic phrases. Firstly, we extract melodic patterns from a collection of audio recordings in an unsupervised way. Next, we group similar patterns by exploiting complex networks concepts and techniques. Drawing an analogy to topic modeling in text classification, we then represent audio recordings using a vector space model. Finally, we employ a number of classification strategies to build a predictive model for raga recognition. To evaluate our approach, we compile a music collection of over 124 hours, comprising 480 recordings and 40 ragas. We obtain 70% accuracy with the full 40-raga collection, and up to 92% accuracy with its 10-raga subset. We show that phrase-based raga recognition is a successful strategy, on par with the state of the art, and sometimes outperforms it.
Introduction
In Indian Art Music (IAM), a rāga is more than just a scale; it is a complex melodic framework characterized by specific notes (svaras), ascending/descending patterns, and, crucially, characteristic melodic motifs or "catch phrases." While seasoned listeners identify rāgas by recognizing these phrases, most computational approaches rely on Pitch Class Distributions (PCD)—essentially histograms of how often each note is played. These "bag-of-notes" methods ignore the temporal "grammar" of the music. This research addresses the challenge of automatically discovering and utilizing these melodic phrases for rāga identification without human intervention, moving beyond manual annotation which is subjective and difficult to scale.
Methodology
The proposed pipeline treats music like text, employing a multi-stage process to transform raw audio into a classification-ready feature matrix.
- Melodic Pattern Discovery: Using an unsupervised system, 2-second melodic segments are extracted from audio recordings.
- Melodic Pattern Clustering: Segments are represented as nodes in a complex network. An optimal similarity threshold is determined by comparing the clustering coefficient of the music network against a randomized version. Community detection algorithms then group similar patterns into clusters, where each cluster represents a unique melodic phrase.
- Vector Space Modeling: Drawing an analogy to Natural Language Processing (NLP), a recording is treated as a document and a melodic phrase as a word. This allows for the construction of a "vocabulary" of phrases.
- Feature Extraction: Three feature types were tested: binary presence ($F_1$), term frequency ($F_2$), and inverse document frequency ($F_3$).
Datasets
The study introduces a significant new resource for the community:
- DB40raga: A diverse dataset of 124 hours of commercially available Carnatic music, featuring 480 recordings across 40 different rāgas.
- DB10raga: A subset of 10 rāgas used for benchmarking against previous studies.
Key Results and Insights
The system achieved high performance across both datasets, with Multinomial Naive Bayes (NBM) emerging as the top-performing classifier. As shown in the results table, the phrase-based approach ($M$) significantly outperformed existing pitch-parameterized methods ($S_2$) and remained competitive with standard pitch-distribution methods ($S_1$).
- The "Catch Phrase" Hypothesis: Results showed that binary presence ($F_1$) was nearly as effective as frequency ($F_2$), confirming that a single "catch phrase" is often sufficient for rāga identification.
- NLP Analogies: The success of NBM suggests that the structural "unfolding" of a rāga in performance is computationally similar to how topics are discussed in text documents.
- Interpretability: Unlike "black box" models, this approach produces a dictionary of discovered melodic phrases that can be listened to and verified by musicians, making the results musically meaningful.
Conclusion
While the system excels at motif-based identification, it faces challenges with "allied rāgas" that share identical notes and similar movements. Future work will focus on combining phrase-based features with pitch-distribution features to create a hybrid model capable of handling the diverse landscape of Indian Art Music with even greater robustness.
Summary updated