TLDR: ShrutiSense is a new symbolic pitch processing system designed for Indian classical music. It addresses the challenges of microtonal distinctions (22 Shrutis) and complex raga grammars, which are often lost with Western 12-tone systems. The system uses a Shruti-aware Finite-State Transducer (FST) for correcting corrupted pitch sequences and a Grammar-Constrained Shruti Hidden Markov Model (GC-SHMM) for completing melodic sequences. The FST model shows high accuracy (91.3%) for pitch correction, making it ideal for real-world applications like digital music education and archival, while the GC-SHMM is more effective for melodic completion.
Indian classical music is renowned for its intricate microtonal system, featuring 22 distinct pitch intervals known as Shrutis within each octave. This system allows for a level of expressive detail and melodic ornamentation far beyond the 12-tone equal temperament found in Western music. However, this unique complexity also poses significant challenges for modern digital music technologies, which often fail to accurately capture these microtonal distinctions and the specific grammatical rules that govern melodic movement within different ragas.
Existing digital tools frequently misinterpret or simplify pitches in Indian classical music, leading to inaccuracies that violate the traditional raga grammar. This problem affects various real-world applications, from digital music education platforms that need to provide precise pitch feedback to students, to archival efforts that aim to preserve historical recordings with their cultural nuances intact. Performance assistance technologies and music information retrieval systems also struggle to analyze and process this music authentically.
To address these critical issues, researchers Rajarshi Ghosh and Jayanth Athipatla have developed ShrutiSense, a comprehensive symbolic pitch processing system specifically designed for Indian classical music. ShrutiSense tackles two primary tasks: correcting westernized or corrupted pitch sequences to align with the 22-Shruti framework and raga grammar, and completing melodic sequences where pitches are missing.
ShrutiSense employs two complementary models for these tasks. For pitch correction, it uses a Shruti-aware Finite-State Transducer (FST). This model excels at making contextual corrections within the 22-Shruti system, ensuring that the output adheres to the specific melodic rules of a raga. For completing melodic sequences with missing values, ShrutiSense utilizes a Grammar-Constrained Shruti Hidden Markov Model (GC-SHMM), which incorporates raga-specific transition rules to infer and fill in the missing pitches based on musical context.
The system formalizes the 22-Shruti system within a computational framework, explicitly encoding raga grammar rules such as ascending (arohana) and descending (avarohana) scales, and permissible transitions between Shrutis. This allows ShrutiSense to balance pitch accuracy with cultural authenticity, ensuring that the corrected or completed music remains true to its traditional form.
In terms of performance, ShrutiSense has shown impressive results. For the crucial pitch correction task, the FST model achieved a remarkable 91.3% Shruti classification accuracy, significantly outperforming other methods and demonstrating robust performance even with considerable pitch noise. This makes the FST pipeline the preferred choice for real-world correction applications, such as processing audio files to transform them into a culturally appropriate Carnatic form.
While the correction task is where ShrutiSense truly shines for practical use, its melodic completion capabilities are also noteworthy. Although generally more challenging, the GC-SHMM model showed better overall accuracy for completion tasks, particularly when dealing with structured missing patterns. The system’s ability to maintain high accuracy across different ragas, such as Yaman, Bhairavi, Bilaval, Kalyan, and Khamaaj, further validates its generalizability.
Also Read:
- Bridging the Gap: DeformTune’s Intuitive AI Music for Novices
- DRKF: Advancing Emotion Recognition Through Decoupled Representations and Knowledge Fusion
Despite its promising advancements, the developers acknowledge certain limitations. These include challenges with rapid microtonal ornamentations (like gamak and meend), cross-raga modulation, and extreme pitch deviations. Future work aims to address these by focusing on adaptive raga learning, improved pitch estimation for noisy audio, more sophisticated ornament modeling, and extensions to handle multi-voice textures common in Indian classical ensembles. ShrutiSense represents a significant step forward in preserving and processing the rich microtonal heritage of Indian classical music in the digital age.
