Counting Still Wins: A Label-Free Evaluation of Melodic Representations for Carnatic Raga Retrieval

Raga retrieval, finding other renditions of the same raga by melodic content, is a core organising task for archives of Indian art music, where expert annotation is scarce and does not scale. We present a label-free evaluation of melodic representations for query-by-example raga retrieval on the Saraga Carnatic collection (176 labelled recordings, 95 ragas, 123 evaluable queries). From tonic-normalised, symbolically denoised svara sequences we compare seven representations spanning the sparse-to-dense axis of modern retrieval: a pitch-class histogram, a first-order svara-transition matrix, BM25, an LSTM language model, a Transformer language model, contrastive fine-tuning with InfoNCE, and ColBERT-style late interaction. Count-based representations (P@1 = 0.423 and 0.431) outperform every learned alternative (best learned: 0.369 with a standard deviation of 0.016 over five held-out-artist splits), in aggregate and within the allied-raga stratum where scale content alone cannot discriminate. We further show that apparent single-run parity of contrastive embeddings is not reproduced under multi-seed evaluation; that late interaction, which removes the pooling bottleneck entirely, performs worse than the pooled model it was built from, eliminating pooling as an explanation; that BM25 retrieves at chance because every pitch position occurs in every recording, so that document frequency carries no information and every query reduces to the same twelve terms; and that symbolic denoising improves retrieval while raising language-model perplexity. A sub-study on the phrase-annotated subset reverses the picture: a term-weighted vector space model over expert phrases outperforms svara counting on the same recordings, and improves further under approximate phrase matching. We therefore state the finding as a claim about representational granularity rather than about model families: at this data scale, the unit of representation matters more than the sophistication of the model.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-14
DOI
https://doi.org/10.5281/zenodo.24229236
Primary Topic
Music and Audio Processing
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Counting Still Wins: A Label-Free Evaluation of Melodic Representations for Carnatic Raga Retrieval

Saanchita V Iyer
Zenodo (CERN European Organization for Nuclear Research)
Music and Audio Processing
preprint

Counting Still Wins: A Label-Free Evaluation of Melodic Representations for Carnatic Raga Retrieval

Saanchita V Iyer
preprint en

Abstract

Raga retrieval, finding other renditions of the same raga by melodic content, is a core organising task for archives of Indian art music, where expert annotation is scarce and does not scale. We present a label-free evaluation of melodic representations for query-by-example raga retrieval on the Saraga Carnatic collection (176 labelled recordings, 95 ragas, 123 evaluable queries). From tonic-normalised, symbolically denoised svara sequences we compare seven representations spanning the sparse-to-dense axis of modern retrieval: a pitch-class histogram, a first-order svara-transition matrix, BM25, an LSTM language model, a Transformer language model, contrastive fine-tuning with InfoNCE, and ColBERT-style late interaction. Count-based representations (P@1 = 0.423 and 0.431) outperform every learned alternative (best learned: 0.369 with a standard deviation of 0.016 over five held-out-artist splits), in aggregate and within the allied-raga stratum where scale content alone cannot discriminate. We further show that apparent single-run parity of contrastive embeddings is not reproduced under multi-seed evaluation; that late interaction, which removes the pooling bottleneck entirely, performs worse than the pooled model it was built from, eliminating pooling as an explanation; that BM25 retrieves at chance because every pitch position occurs in every recording, so that document frequency carries no information and every query reduces to the same twelve terms; and that symbolic denoising improves retrieval while raising language-model perplexity. A sub-study on the phrase-annotated subset reverses the picture: a term-weighted vector space model over expert phrases outperforms svara counting on the same recordings, and improves further under approximate phrase matching. We therefore state the finding as a claim about representational granularity rather than about model families: at this data scale, the unit of representation matters more than the sophistication of the model.

Zenodo (CERN European Organization for Nuclear Research)
Music and Audio Processing
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon