Abstract
We perform a comparative evaluation of methodologies for computing similarity between short-time melodic fragments of audio recordings of Indian art music. We experiment with 560 different combinations of procedures and parameter values. These include the choices made for the sampling rate of the melody representation, pitch quantization levels, normalization techniques and distance measures. The dataset used for evaluation consists of 157 and 340 annotated melodic fragments of Carnatic and Hindustani music recordings, respectively. Our results indicate that melodic fragment similarity is particularly sensitive to distance measures and normalization techniques. Sampling rates do not have a significant impact for Hindustani music, but can significantly degrade the performance for Carnatic music. Overall, the performed evaluation provides a better understanding of the processing steps and parameter settings for melodic similarity in Indian art music. Importantly, it paves the way for developing unsupervised melodic pattern discovery approaches, whose evaluation is a challenging and, many times, ill-defined task.
Introduction
Indian Art Music (IAM) is primarily an oral tradition where melodies are governed by the framework of a rāg. Within a rāg, certain melodic motifs or "catch phrases" (often called pakads) define its identity. Automatically detecting these motifs is essential for music retrieval, recommendation, and transcription. However, computing melodic similarity is notoriously difficult in IAM due to creative improvisation, varying artist tonics (transposition), and non-linear temporal variance.
While many algorithms exist for melodic similarity, there has been no consensus on which specific processing steps—like sampling rate, normalization, or distance metrics—actually work best for the unique characteristics of IAM. This study performs a massive "grid search" evaluation of 560 different algorithmic combinations to establish these benchmarks.
Methodology
The authors systematically evaluated 560 variants of a similarity pipeline consisting of four main stages:
- Melody Representation: Extracting the predominant pitch (in Cents) at five sampling rates ranging from 33 Hz to 100 Hz.
- Transposition Invariance (Normalization): Testing eight techniques to handle pitch shifts, including Tonic normalization (using the artist's base frequency), Zero-mean normalization, and Z-normalization.
- Uniform Time-Scaling: Comparing raw fragments against generating five time-stretched versions (0.9x to 1.1x) of each fragment to handle global tempo changes.
- Similarity Computation: Comparing standard Euclidean distance against seven variants of Dynamic Time Warping (DTW), experimenting with different local and global constraints.
Datasets
The study utilized two distinct datasets to account for the major traditions of IAM:
- Carnatic Music Dataset (CMD): 23 recordings, featuring 157 annotated melodic fragments across 5 pattern types.
- Hindustani Music Dataset (HMD): 9 recordings, featuring 340 annotated melodic fragments across 5 pattern types. To simulate real-world retrieval tasks, thousands of random melodic segments were injected into the search space as noise.
Key Results and Discussion
The performance was measured using Mean Average Precision (MAP). The results indicate that optimal settings are tradition-specific, largely due to the differing ornamental styles of Hindustani and Carnatic music.
The "Gamaka" Factor and Sampling Rates
Carnatic music relies heavily on fast, oscillatory melodic ornaments known as gamakas. Consequently, the study found that Carnatic music requires higher sampling rates (67-100 Hz) to maintain accuracy. In contrast, Hindustani music, which often features smoother transitions, was much more robust to lower sampling rates.
Normalization and Distance Measures
For Hindustani music, normalizing pitch relative to the artist's tonic was most effective. However, in Carnatic music, Zero-mean normalization performed better. This is likely because Carnatic motifs often recur in different pitch regions (tetra-chords) within the same performance, which tonic normalization does not account for. Across the board, DTW-based measures significantly outperformed Euclidean distance, confirming that timing variations in IAM are non-linear.
As shown in the results table, Hindustani music (HMD) achieved a top MAP of 0.552, while Carnatic music (CMD) reached 0.413. Statistical significance tests confirmed that the differences between these 560 variants were not due to random noise, emphasizing the importance of parameter selection.
Conclusion
This evaluation provides a better understanding of the processing steps required for melodic similarity in IAM. The findings highlight that while DTW is a superior distance measure, normalization and sampling rates must be tailored to the specific musical tradition. Future work will focus on developing better unsupervised melodic segmentation and incorporating additional features like loudness and timbre into similarity measures.
Summary updated