Tachyeon/audio-fingerprint-indian-bench

Audio Fingerprinting Benchmark on Indian Classical Music

A reproducible, pre-registered benchmark of five audio-fingerprinting systems on the Saraga 1.5 corpus (Hindustani + Carnatic), plus a pre-registered training-recipe improvement to the NAFP baseline that achieves Bonferroni-significant gains on 1-second queries.

v0.7  ·  closed-world retrieval  ·  5 systems  ·  6 528 evaluation cells  ·  pooled McNemar p = 3.18 × 10⁻⁶

---

TL;DR

• 5 systems benchmarked: Olaf, Dejavu, Panako (classical hash-based); NAFP (Chang et al. ICASSP 2021); NMFP (Araz et al. ISMIR 2025)
• 357 reference tracks (108 Hindustani + 249 Carnatic) from Saraga 1.5 — clean library, no distractors (FMA mix-in out of scope)
• 1 632 queries × 4 lengths (1 s / 3 s / 5 s / 10 s) = 6 528 evaluation cells per system
• Pre-registered improvement (recipe v3): NAFP HR@1 on main_1s improves from 0.983 → 0.995 mean across 3 seeds, pooled McNemar p = 3.18 × 10⁻⁶ (Bonferroni-significant at α/8 = 0.00625)
• ~68 % fewer misses across the benchmark (~71 % on the hardest cell)
• Two pre-registered negative results with identical rigor: (1) per-artist mean subtraction at inference and (2) hubness post-processing (Inverted Softmax + CSLS) — strengthen the recipe v3 positive
• Sample-accurate ground truth: every query stores exact (ref_id, offset_sec, seed) — alignment error measurable to sub-50 ms
• Source MP3s not redistributed: rebuild library by fetching Saraga 1.5 directly from Zenodo
• Full primary endpoint writeup: data/results/RESULTS_recipe_v3.md; pre-registration: data/results/PROTOCOL_recipe_v3.md

---

Headline result

HR@1 — main set (1 000 queries, no section constraint)

System	1 s	3 s	5 s	10 s
Olaf	0.492	0.955	0.993	0.998
Dejavu	0.745	0.969	0.994	1.000
Panako	0.000	0.000	0.922	0.997
NAFP (Chang et al. 2021 — our baseline)	0.983	0.998	0.999	1.000
Recipe v3 (3-seed mean — this work)	0.995	1.000	1.000	1.000
NMFP-ckpt-100 (Araz et al. 2025 — ceiling)	1.000	1.000	1.000	1.000

HR@1 — ablation set (632 section-aligned queries)

System	1 s	3 s	5 s	10 s
Olaf	0.494	0.907	0.987	1.000
Dejavu	0.764	0.978	0.998	1.000
Panako	0.000	0.000	0.929	0.994
NAFP (baseline)	0.979	1.000	1.000	1.000
Recipe v3 (3-seed mean)	0.991	1.000	1.000	1.000
NMFP-ckpt-100 (ceiling)	1.000	1.000	1.000	1.000

Statistical significance (recipe v3 vs baseline, pooled across 3 seeds)

Cell	Baseline	Recipe v3 (mean)	Pooled b, c	Pooled McNemar p
main 1 s	0.983 (17 miss)	0.995 (5 miss)	b = 12, c = 48	3.18 × 10⁻⁶ ✓
main 3 s	0.998	1.000	b = 0, c = 6	0.031
main 5 s	0.999	1.000	b = 0, c = 3	0.250
main 10 s	1.000	1.000	—	—
ablation 1 s	0.979 (13 miss)	0.991 (5.7 miss)	b = 17, c = 39	0.0046 ✓
ablation 3 s / 5 s / 10 s	1.000	1.000	—	—

✓ Bonferroni-significant at α/8 = 0.00625. The two hardest cells clear the corrected threshold; remaining six cells are at ceiling (no statistical headroom). Pooled CSV: data/results/recipe_v3_pooled_mcnemar.csv.

Error-rate reduction: main_1s −70.6 % misses (17 → 5), ablation_1s −56.2 % (13 → 5.7), benchmark-wide −67.6 % (33 → 10.7 across 6 528 cells).

---

What's in this dataset

data/
├── queries/                                  # AudioFolder
│   ├── metadata.parquet                      # 1 000 main queries with ground-truth offsets
│   ├── hindustani/*.wav                      # 500
│   └── carnatic/*.wav                        # 500
├── queries_ablation/
│   ├── metadata.parquet                      # 632 section-aligned queries
│   ├── hindustani/*.wav                      # 207
│   └── carnatic/*.wav                        # 425
├── refs.parquet                              # 357 ref tracks; metadata only (no audio paths)
├── results/
│   ├── {system}_{split}_{length}.parquet     # ranked top-K candidates per query
│   ├── {system}_{split}_{length}.scores.json # HR@k + Wilson 95 % CI + alignment error
│   ├── recipe_v3_seed{42,137,2026}_{split}_{length}.parquet
│   ├── recipe_v3_seed{42,137,2026}_{split}_{length}.scores.json
│   ├── recipe_v3_pooled_mcnemar.csv          # primary endpoint, pooled across 3 seeds
│   ├── PROTOCOL_recipe_v3.md                 # pre-registration (locked before training)
│   ├── PROTOCOL_intervention2.md             # pre-registered NEGATIVE result protocol
│   └── RESULTS_recipe_v3.md                  # full writeup of the primary endpoint
├── inspection/                               # Library audit tables
│   ├── tracks.parquet                        # 357 rows: ref_id, corpus, raagas, taalas, artists, works, work_mbids
│   ├── sections.parquet                      # section_type spans per ref (alaap / composed / tani)
│   ├── works.parquet                         # unique work-MBIDs
│   └── leakage_pairs.parquet                 # 130 (main_query → other_ref) pairs sharing a work-MBID
└── configs/                                  # Seeded test-set generation configs (reproducibility)
    ├── hindustani_main.json
    ├── hindustani_ablation.json
    ├── carnatic_main.json
    └── carnatic_ablation.json

ARTEFACTS.parquet                             # SHA-256 + size manifest for every file in this dataset
ATTRIBUTION.md                                # full upstream credits (Saraga / NAFP / NMFP / FMA)
LICENSE        / LICENSE-CODE                 # CC-BY-NC-SA 4.0 (data) / MIT (code in linked repo)

{system} ∈ {olaf, dejavu, panako, nafp, nmfp} · {split} ∈ {main, ablation} · {length} ∈ {1s, 3s, 5s, 10s}.

---

How to load

Audio queries — registered as HF datasets configs:

from datasets import load_dataset

# 1 000 main queries with sample-accurate ground-truth offsets
queries = load_dataset(
    "Tachyeon/audio-fingerprint-indian-bench",
    "queries",
    split="test",
)
# queries[0] → {'audio': {...}, 'query_id': ..., 'ref_id': ..., 'offset_sec': ...}

# 632 section-aligned ablation queries
ablation = load_dataset(
    "Tachyeon/audio-fingerprint-indian-bench",
    "queries_ablation",
    split="test",
)

Everything else (parquets / scores / protocols) — fetched directly with hf_hub_download. The full file tree is browsable on the dataset's Files tab:

from huggingface_hub import hf_hub_download
import pandas as pd, json

# 357-row reference-track metadata (no audio paths — fetch source MP3s from Zenodo separately)
refs_path = hf_hub_download(
    "Tachyeon/audio-fingerprint-indian-bench",
    "data/refs.parquet", repo_type="dataset",
)
refs = pd.read_parquet(refs_path)        # 357 rows

# Any (system × split × length) score, e.g. recipe v3 seed 42 on main_1s
scores_path = hf_hub_download(
    "Tachyeon/audio-fingerprint-indian-bench",
    "data/results/recipe_v3_seed42_main_1s.scores.json",
    repo_type="dataset",
)
print(json.load(open(scores_path))["hr@1"])   # → 0.993

# The 130 composition-twin pairs (library audit)
twins = pd.read_parquet(hf_hub_download(
    "Tachyeon/audio-fingerprint-indian-bench",
    "data/inspection/leakage_pairs.parquet",
    repo_type="dataset",
))

---

Methodology

Query construction

• 10-second clips at 16 kHz mono, cut at random offsets from Saraga ref tracks
• Sample-accurate: each query stores (ref_id, offset_sec, seed) — ground-truth alignment is bit-exact
• Main set (1 000 queries): random tracks + random offsets, no section constraint (500 Hindustani + 500 Carnatic)
• Ablation set (632 queries): offsets aligned to Saraga's section_annotation metadata; section_type ∈ {alaap, composed, tani} enables per-section breakdowns
• Shorter (1/3/5 s) queries are first-N truncations of the 10-second cuts — not random re-cuts. Documented as a methodological limitation; see Limitations.

Reference library

• 357 Saraga 1.5 concert recordings; no other audio added (closed-world, clean library)
• 130 main queries share a MusicBrainz work-MBID with a different recording in the library (composition-twin leakage). Tracked explicitly via data/inspection/leakage_pairs.parquet; HR@1 reported separately for with-twin / no-twin subsets inside each scores.json.

Scoring

• HR@k: fraction of queries where the truth ref_id is among the top-k predicted refs. Reported for k ∈ {1, 5, 10} with Wilson 95 % CI.
• MRR@k: mean reciprocal rank, capped at k.
• top1_near: HR@1 with a coarse temporal-accuracy gate (within ±0.5 s of the truth offset); follows the NAFP-paper convention.
• Alignment error: per-query offset error (median, p95, max), reported in seconds.

Statistical comparison

• McNemar exact test on paired binary (hit / miss) outcomes per query
• 3-seed pooled McNemar for the recipe v3 primary endpoint (3 000 paired observations per cell)
• Bonferroni correction at α/8 = 0.00625 across the 8 (split × length) cells

---

Systems benchmarked

System	Type	Training	Reference
Olaf	Classical (constellation hash)	None (rule-based)	https://github.com/JorenSix/Olaf
Dejavu	Classical (peak pairs, PostgreSQL)	None	https://github.com/worldveil/dejavu
Panako	Classical (CQT triplet hash)	None	http://panako.be
NAFP	Neural CNN + NT-Xent contrastive	10 epochs on FMA-medium 10k_icassp	Chang et al. ICASSP 2021
NMFP	Neural CNN (same architecture as NAFP)	100 epochs on FMA-medium with 5 recipe fixes	Araz et al. ISMIR 2025

NMFP weights are pre-trained by Araz et al. (Zenodo 15719945, GPLv3 / AGPLv3 — viral); we use them only to establish the ceiling and do not redistribute.

---

Recipe v3 — pre-registered training-recipe improvement (this work)

Hypothesis (locked in PROTOCOL_recipe_v3.md before training): two of NMFP's published recipe fixes, combined with a larger batch and a 3× longer schedule, will recover NAFP's same-artist failure mode on 1-second queries with Bonferroni-significant gain across 8 cells × 3 seeds.

Recipe: F_MIN: 300 → 160 Hz · TR_SEG_MODE: all → random_oneshot (one anchor per track per epoch, resampled) · BSZ: 120 → 320 · MAX_EPOCH: 10 → 30 · NT-Xent τ = 0.05 preserved · Adam, cosine LR. Trained from scratch on FMA-medium (same training corpus as baseline). 80.3 min total wall across 3 seeds on a Colab L4 GPU.

Result: primary endpoint cleared — main_1s pooled McNemar p = 3.18 × 10⁻⁶ (clears Bonferroni α/8 by ~1 900×). ablation_1s pooled p = 0.0046. No cell regresses on average.

What we did NOT do (honest disclosure): we did not apply NMFP's other 3 recipe fixes — full-IR augmentation, 1-second acoustic history, triplet loss with semi-hard mining — because each requires dataloader/model surgery beyond our budget. Adding them is the path to closing the residual ~0.005 gap to NMFP's 1.000 ceiling.

---

Pre-registered negative results

Every intervention's hypothesis, primary endpoint, and falsification rule are committed before training data is collected. Two interventions tested with this protocol returned negative.

Intervention 2 — per-artist mean subtraction at inference

• Hypothesis: subtract a learned per-artist centroid from each NAFP ref embedding to reduce same-artist top-1 confusions
• Design: α-sweep {0, 0.05, 0.10, 0.15, 0.20, 0.25, 0.30} + 3 controls — α = 0 sanity, shuffled-centroid (artist labels permuted), and isotropic-centroid (random unit-norm vector replacing the learned centroid)
• Result: REJECTED. The isotropic-centroid control matched the baseline in 4 / 4 main cells — the mechanism is not artist-specific. Same-artist confusion is not addressable by inference-time centroid correction.
• Protocol: data/results/PROTOCOL_intervention2.md

Hubness post-processing — Inverted Softmax + CSLS

• Hypothesis: NAFP's same-artist failure mode is a generic hubness problem fixable at inference time via Smith et al. 2017 / Conneau et al. 2018 re-scoring
• Result: REJECTED. Both methods reproduced baseline HR@1 exactly across all 8 cells (no gain, no regression). The failure is encoder-level, not embedding-geometry.

Both negatives strengthen the recipe v3 positive: inference-only fixes were ruled out, isolating training-recipe modifications as the actual lever.

---

Reproducibility

# 1. Fetch dataset metadata + queries + result parquets from this dataset
from datasets import load_dataset
queries = load_dataset("Tachyeon/audio-fingerprint-indian-bench", "queries", split="test")

# 2. Fetch Saraga 1.5 source MP3s (NOT redistributed here; required for reference library)
#    pip install mirdata
#    import mirdata
#    mirdata.initialize("saraga_hindustani", data_home="~/saraga_hindustani").download()
#    mirdata.initialize("saraga_carnatic",   data_home="~/saraga_carnatic").download()

# 3. Code, system runners, and the recipe v3 training pipeline live in a separate repo:
#    https://github.com/ipritamdash/afp-indian-classical  (private; request access via citation email)

# 4. NMFP-ckpt-100 weights (optional — only needed to reproduce the ceiling row):
#    https://zenodo.org/records/15719945   (GPLv3 / AGPLv3; review before bundling into derivative work)

Per-query result parquets are deterministic given fixed (system, seed, audio). Recipe v3 seeds: 42, 137, 2026 (locked in PROTOCOL_recipe_v3.md before training).

---

Limitations

• Small reference library (357 refs). HR@1 saturates for ≥ 3-second queries on all neural systems. The interesting signal lives in the 1-second cells; deployment-scale benchmarks would mix in 10 k+ distractors (out of scope here).
• Shorter queries are first-N truncations of the 10-second cuts — not random re-cuts. Inherited from the Chang et al. 2021 protocol; documented but not fixed.
• Recipe v3 trades epoch parity for convergence — trained 30 epochs vs the baseline's 10 epochs. The improvement is recipe + 3× longer training combined; per-knob ablations were out of scope.
• 3 seeds is the minimum for a stable pooled-McNemar claim; 5+ seeds is preferable in follow-up.
• NMFP-ckpt-100 remains the ceiling at HR@1 = 1.000 across all 8 cells. We reach ~95 % of that ceiling at ~10 % of their training compute, but do not beat it.
• Tani per-section cell is underpowered (N = 11 in ablation_1s; Carnatic-only — Hindustani has no tani sections by tradition). Per-section HR@1 claims for tani should be treated as descriptive only.

---

Licensing

Component	License
Reference-track metadata, query metadata, inspection tables	CC-BY-NC-SA 4.0 (inherits Saraga)
Query WAVs (derived from Saraga source)	CC-BY-NC-SA 4.0
Per-system result parquets, scores.json	CC-BY-NC-SA 4.0 (covers derivative-work definition)
Source MP3s	NOT distributed — fetch from Zenodo (CC-BY-NC-SA 4.0)
NMFP teacher weights	NOT distributed — fetch from Zenodo 15719945 (GPLv3 / AGPLv3)
Code (separate GitHub repo)	MIT — applies only to code in the linked repo, not to data here

See LICENSE, LICENSE-CODE, and ATTRIBUTION.md for full text.

---

Citation

@misc{banwala2026afpindianclassical,
  author       = {Aryan Banwala},
  title        = {Audio Fingerprinting Benchmark on Indian Classical Music},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Tachyeon/audio-fingerprint-indian-bench}},
  note         = {Includes a pre-registered training-recipe improvement to NAFP (Chang et al. 2021), Bonferroni-significant on 1-second queries.},
}

When citing this benchmark, please also cite the upstream papers:

@dataset{srinivasamurthy2021saraga,
  title     = {{Saraga}: Open Datasets for Research on {I}ndian Art Music},
  author    = {Srinivasamurthy, Ajay and Gulati, Sankalp and Repetto, Rafael Caro and Serra, Xavier},
  year      = {2021},
  version   = {1.5},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.4301737},
}
@inproceedings{chang2021nafp,
  title     = {{Neural Audio Fingerprint} for High-specific Audio Retrieval based on Contrastive Learning},
  author    = {Chang, Sungkyun and Lee, Donmoon and Park, Jeongsoo and Lim, Hyungui and Lee, Kyogu and Ko, Karam and Han, Yoonchang},
  booktitle = {ICASSP},
  year      = {2021},
  doi       = {10.1109/ICASSP39728.2021.9414083},
}
@inproceedings{araz2025nmfp,
  title     = {Enhancing Neural Audio Fingerprint Robustness to Real-World Conditions},
  author    = {Araz, R. O. and Cortès-Sebastià, G. and Molina, E. and Serra, X. and Serra, J. and Mitsufuji, Y. and Bogdanov, D.},
  booktitle = {ISMIR},
  year      = {2025},
  eprint    = {arXiv:2506.22661},
}

---

Version

v0.7  ·  2026-05-17  ·  current and only supported release.

---

Attribution

See ATTRIBUTION.md for full credits to Saraga / NAFP / NMFP / FMA upstream authors.