Computational Raga Recognition: A Theory-Driven Approach

Janay Sukheja, Shailaja D. SharmaarXiv 2026 · arXiv

TL;DR

This paper introduces a theory-driven computational framework for recognizing Hindustani classical ragas by mapping musicological concepts—like resting notes (Nyasa) and characteristic phrases (Chalan)—directly to audio features. By utilizing a high-resolution 44-position pitch framework based on the traditional 22-shruti system, the researchers successfully distinguished Raga Yaman from other ragas with an 84% similarity score, demonstrating that explicit musicological rules can be as effective as "black-box" machine learning models.

What Problem Does This Paper Address?

In Hindustani classical music, a raga is far more than just a scale or a set of notes. Two ragas might share the exact same notes but remain distinct based on which notes are emphasized, how a performer rests on a note, and the specific melodic pathways (Chalan) used to move between them.

Traditional computational approaches often rely on global pitch distributions or generic machine learning features that lose these subtle, time-dependent musicological nuances. The authors aim to bridge the gap between formal music theory and signal processing, creating a model where every computational feature corresponds to a specific concept taught to human students of the art form.

Methods

The authors developed a multi-stage pipeline to extract and compare raga characteristics:

  1. Pitch Representation: After isolating vocals and extracting pitch using the pYIN algorithm, the system maps frequencies to a 44-position note framework. This is an extension of the traditional 22-shruti system, using a 25%-50%-25% partitioning between shrutis to capture the fine-grained pitch movements and ornamentations typical of Indian classical music.

    Fig. 2. Visual representation of intermediate note
  2. Feature Extraction: The model extracts six core musicological features:

  3. Nyasa: Resting notes held for at least one second.
  4. Purvang-Uttarang: The balance of activity between the lower and upper halves of the octave.
  5. Vadi-Samvadi: The primary and secondary dominant notes.
  6. Alpatva/Varjit: Notes that are used sparingly or avoided entirely.
  7. Aarohan-Avrohan: The specific patterns of ascending and descending transitions.
  8. Chalan: Recurring melodic phrases, analyzed using a weighted Levenshtein distance to allow for slight performance variations.
  9. Similarity Scoring: A reference model was built using recordings of Raga Yaman. New recordings are compared against this reference using feature-specific metrics (like angular distance for distributions) to generate an overall similarity score.
Fig. 1. Complete methodological framework

Datasets

The study utilized a curated dataset of Hindustani classical vocal performances:

  • Reference Set: 11 audio recordings of Raga Yaman, totaling approximately 6 hours of performance by expert musicians.
  • Test Set: Recordings of Yaman as well as five other ragas for comparison: Shuddha Kalyan, Bhoop, Bageshri, Miyan Malhar, and Marwa.
  • Preprocessing: Lead vocals were isolated using RipX DAW to minimize interference from accompaniment instruments like the harmonium or sarangi.

Key Insights and Ideas

  • The Power of "Small Data": Unlike deep learning models that require thousands of samples, this theory-driven approach achieved high specificity using only 11 reference recordings because it "knows" what to look for based on centuries of musicological tradition.
  • Resolution Matters: The 44-position framework is vital. Standard 12-tone Western scales are too coarse to capture the "slides" (meend) and microtonal nuances that define a raga's identity.
  • Tempo Sensitivity: The model performs best on Alap (slow, rhythm-free introduction) or Vilambit (slow tempo) sections. In these sections, notes are sustained long enough for the system to reliably identify Nyasa (resting) swaras.
  • Musicological Overlap: The model gave Raga Shuddha Kalyan a relatively high similarity score (64.9%) to Yaman. This isn't a failure; rather, it accurately reflects the fact that these two ragas share significant theoretical characteristics.

Key Results

The framework was highly effective at identifying the target raga while distinguishing it from others:

  • Raga Yaman (Target): Achieved a similarity score of 84.0%.
  • Shuddha Kalyan: Scored 64.9%, reflecting its close theoretical relationship to Yaman.
  • Distinct Ragas: Bageshri (18.9%), Miyan Malhar (13.5%), and Marwa (0.0%) scored significantly lower, showing the model's ability to reject ragas with different structures.

Analysis of the reference recordings showed that Sa and Pa were the most frequent resting notes (Nyasa) for Yaman, as shown in the distribution below:

Fig. 3. Nyasaswara distribution in reference recording

Limitations and Future Work

  • Tonic Dependency: The system currently requires the manual identification of the tonic (Sa). Automated tonic estimation is a necessary next step for a fully autonomous system.
  • Contextual Blindness: The current Vadi-Samvadi analysis relies on note prominence but doesn't yet account for the melodic context (e.g., a note's position at the end of a specific phrase).
  • Performance Variation: The model struggles with Drut (fast) passages where notes are not sustained long enough to meet the 1-second Nyasa stability criterion.
  • Scalability: While successful for Yaman, the framework needs to be tested across a much wider and more diverse set of ragas to prove its general applicability.

Summary updated

Abstract

Raga recognition in Hindustani classical music is difficult because a raga is not defined only by the notes used in it. Its identity also depends on the way notes are emphasized, the movement between notes and characteristic phrases. In this work, we develop a computational approach to analyse these characteristics directly from audio recordings. The method uses pitch extracted from recordings and represents it using a 44-position note framework based on the 22-shruti system. From this representation, Raga features based on the Indian classical musicology are extracted.Each feature is compared with a reference representation of the raga and the resulting scores are combined to obtain an overall similarity score. The method is first studied using Raga Yaman and is then tested on recordings of other ragas. The computational results are also compared with the theoretical description of Yaman to identify which raga characteristics are captured well and where the method has limitations.
BibTeX
@misc{sukheja2026computational,
  author = {Janay Sukheja and Shailaja D. Sharma},
  title = {{Computational Raga Recognition: A Theory-Driven Approach}},
  year = {2026},
  eprint = {2611.15242},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.15242},
}
Figure