Abstract
Identifying the tonic (Sa) is a fundamental prerequisite for the computational analysis of Indian Classical Music (ICM). Traditional monophonic pitch-tracking methods often fail due to melody bias and octave ambiguity. This paper introduces a multipitch-based pipeline that prioritizes the background drone (the Tanpura) over the lead melody. By combining a harmonic salience function with a C4.5 decision tree classifier, the proposed system achieves an overall identification accuracy of 93% across a diverse dataset of Hindustani and Carnatic recordings, successfully identifying the correct octave for both male and female performers.
Introduction
In Indian Classical Music, the tonic serves as the absolute reference for all melodic and harmonic structures. Unlike Western music, where the tonic is often fixed by the composition, ICM performers choose a tonic that suits their specific vocal or instrumental range. Automated identification of this "North Star" is essential for higher-level tasks such as raga recognition and melodic transcription.
Previous attempts at tonic identification have been hindered by three primary factors: a reliance on lead-melody tracking, the "octave problem" (identifying the pitch class but not the specific register), and the wide variance in tonic frequencies between male and female performers. This study addresses these challenges by treating the audio as a multipitch signal, allowing the system to "hear" the persistent drone even when masked by a louder lead performer.
Methodology
The proposed system follows a four-stage pipeline designed to extract and validate tonic candidates based on musical logic rather than raw signal power.
1. Sinusoid Extraction and Salience
The process begins with a Short-Time Fourier Transform (STFT) to identify raw spectral peaks. A harmonic salience function is then applied to calculate the prominence of various pitches over time. This step is crucial as it emphasizes the steady, reinforcing harmonics of the drone instrument, which typically plays the tonic and the fifth (Pa).
2. Candidate Generation and Selection
A pitch histogram is constructed from a 3-minute audio excerpt, from which the top ten most frequent pitches are extracted as tonic candidates. To resolve the final selection, the system employs a C4.5 decision tree classifier. This classifier evaluates the musical intervals between candidates—for instance, recognizing the characteristic 7-semitone gap between the tonic and the fifth—to distinguish the true Sa from other prominent scale degrees.
Experimental Setup
The system was evaluated using 364 three-minute excerpts from the CompMusic project, covering both Hindustani (38%) and Carnatic (62%) traditions. The dataset included 231 unique performances by 36 artists, with a gender distribution of 80% male and 20% female. Ground truth was established through manual annotation, ensuring that the correct octave was identified for every file.
Results and Discussion
The system achieved a high degree of robustness, with an overall classification accuracy of 93%.
Analysis of the results reveals several key insights:
- Tradition Variance: Accuracy was significantly higher in Hindustani music (98%) compared to Carnatic music (90%). This discrepancy is attributed to the often "weaker" drone presence in Carnatic recordings, which reduces the tonic's prominence in the salience function.
- Gender Performance: The system performed better for male singers (95%) than female singers (88%). In some female performances, the higher-octave tonic fell outside the analysis window, leading to misses.
- Octave Accuracy: By incorporating interval logic, the system successfully identified the tonic across a range of nearly two octaves (110 Hz to 370 Hz), a significant improvement over pitch-class-only models.
Conclusion
By shifting the focus from the lead melody to the multipitch environment of the background drone, this research provides a more reliable foundation for ICM analysis. The integration of musical interval logic via decision trees effectively resolves ambiguities that stumped previous signal-processing-only approaches. Future work will focus on improving performance for female vocalists and recordings with low drone-to-melody ratios through gender-specific models and expanded frequency analysis windows.
Summary updated