Enhancing Neural Audio Fingerprint Robustness to Audio Degradation for Music Identification
Paper • 2506.22661 • Published
Datasets:
Tasks:
Audio Classification
Formats:
soundfolder
Size:
1K - 10K
ArXiv:
Tags:
audio-fingerprinting
music-information-retrieval
indian-classical-music
hindustani
carnatic
saraga
License:
Search is not available for this dataset
No rows found.
Audio Fingerprinting Benchmark on Indian Classical Music
Reproducible 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.
v0.6 (2026-05-14) — see Changelog for what's new.
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
- 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 1-second queries improves from 0.983 → 0.995 mean across 3 seeds (pooled McNemar p = 3.18 × 10⁻⁶)
- Two pre-registered negative results reported with the same rigor (Intervention 2: per-artist mean subtraction; hubness post-processing)
- Sample-accurate ground truth: each query stores exact
offset_secin its source reference recording — alignment error measurable to sub-50 ms - Source MP3s not redistributed: rebuild library by fetching Saraga 1.5 directly from Zenodo
Headline result (HR@1 across 5 systems × 4 query lengths)
Main set (1,000 queries, no section-alignment 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 |
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 (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 (6 miss) | b=17, c=39 | 0.0046 ✓✓ |
| ablation 3 s | 1.000 | 1.000 | — | — |
| ablation 5 s | 1.000 | 1.000 | — | — |
| ablation 10 s | 1.000 | 1.000 | — | — |
✓✓ = Bonferroni-significant at α / 8 = 0.00625. No cell regresses on average. Recipe v3 closes the gap to the NMFP ceiling substantially at ~10 % of NMFP's training compute.
Improvement summary (error-rate reduction)
| Cell | Baseline misses | Recipe v3 mean misses | Error reduction |
|---|---|---|---|
| main 1 s | 17 / 1 000 | 5 / 1 000 | −70.6 % |
| ablation 1 s | 13 / 632 | 5.7 / 632 | −56.2 % |
| main 3 s | 2 / 1 000 | 0 / 1 000 | −100 % |
| main 5 s | 1 / 1 000 | 0 / 1 000 | −100 % |
| Other 4 cells | 0 | 0 | — |
| All 8 cells (total) | 33 / 6 528 | 10.7 / 6 528 | −67.6 % |
Headline: ~68 % fewer misses across the benchmark, ~71 % on the hardest cell.
What's in this dataset
data/
├── queries/ # AudioFolder
│ ├── metadata.parquet # 1 000 main queries × 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/ # Per (system × split × length) parquets + scores.json
│ ├── {system}_{split}_{length}.parquet # ranked top-K candidates per query
│ ├── {system}_{split}_{length}.scores.json # HR@k + Wilson 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 seeds)
│ ├── PROTOCOL_recipe_v3.md # pre-registration (locked before training)
│ ├── PROTOCOL_intervention2.md # pre-registered NEGATIVE result
│ └── RESULTS_recipe_v3.md # full writeup
├── inspection/ # Library audit tables
│ ├── tracks.parquet # 357 rows: ref_id, corpus, raagas, taalas, artists, works, work_mbids
│ ├── sections.parquet # ~2 000 rows: ref_id, start_sec, end_sec, section_type
│ ├── 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
{system} ∈ {olaf, dejavu, panako, nafp, nmfp}. {split} ∈ {main, ablation}. {length} ∈ {1s, 3s, 5s, 10s}.
How to load
from datasets import load_dataset
# Audio queries with sample-accurate offsets
queries = load_dataset("Tachyeon/audio-fingerprint-indian-bench",
"queries", split="test")
# queries[0] → {'audio': {...}, 'query_id': ..., 'ref_id': ..., 'offset_sec': ...}
# Reference-track metadata
refs = load_dataset("Tachyeon/audio-fingerprint-indian-bench",
"refs", split="library")
# Per-system results — load any specific (system × split × length) parquet
import pandas as pd
from huggingface_hub import hf_hub_download
scores_path = hf_hub_download(
repo_id="Tachyeon/audio-fingerprint-indian-bench",
filename="data/results/recipe_v3_seed42_main_1s.scores.json",
repo_type="dataset",
)
import json
print(json.load(open(scores_path))["hr@1"]) # → 0.993
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): no section-alignment constraint
- Ablation set (632): queries align to Saraga's
section_annotationmetadata;section_type∈{alaap, composed, tani}allows 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 limitation; see Limitations.
Reference library
- 357 Saraga 1.5 concert recordings (no other audio added)
- No distractors / no FMA mix — clean retrieval benchmark
- 130 main queries share a MusicBrainz work-MBID with a different recording in
the library (composition-twin leakage). Tracked explicitly via
inspection/leakage_pairs.parquet; we report HR@1 separately for with-twin / no-twin subsets inscores.json.
Scoring
- HR@k: fraction of queries where the truth
ref_idis 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 at coarse temporal accuracy (within ±0.5 s of truth); follows NAFP-paper convention.
- Alignment error: per-query offset error (median, p95, max).
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 arch 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 at data/results/PROTOCOL_recipe_v3.md before training):
two of NMFP's published recipe fixes, combined with a 3× larger batch and 3×
longer schedule, will improve 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; one-anchor-per-track sampler (re-sampled per epoch); 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).
Result: primary endpoint cleared — main_1s pooled p = 3.18 × 10⁻⁶ (clears Bonferroni α / 8 by ~1900×). ablation_1s pooled p = 0.0046 (Bonferroni-sig). 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-sec acoustic history, triplet loss with semi-hard mining) because they require dataloader 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
We use a pre-registration discipline: 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
- α sweep {0, 0.05, …, 0.30} + 3 controls (α=0 sanity, shuffled-centroid, isotropic-centroid)
- Result: REJECTED. Isotropic-centroid control matched V1 in 4/4 main cells → mechanism falsified. The 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 style 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 result: we ruled out inference-only fixes, isolating training-recipe modifications as the actual lever.
Reproducibility
# 1. Fetch dataset metadata + queries + results parquets
from datasets import load_dataset
queries = load_dataset("Tachyeon/audio-fingerprint-indian-bench", "queries", split="test")
# 2. Fetch Saraga 1.5 source MP3s (NOT in this dataset; required for reference library)
# via mirdata:
# 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 + recipe v3 pipeline:
# https://github.com/ipritamdash/afp-indian-classical (currently private; ask author)
# 4. NMFP-ckpt-100 weights (for ceiling reference, optional):
# Zenodo: https://zenodo.org/records/15719945
# License: GPLv3 / AGPLv3 (viral); review before bundling into derivative work
Per-query result parquets are deterministic given fixed (system, seed, audio). Recipe v3 random_seed = 42, 137, 2026.
Limitations
- Library is small (357 refs) — HR@1 saturates for 5/10 s queries on the baseline. The hardest cell (1 s, same-artist) is where the improvement lives.
- No FMA distractors / no expanded gallery — a more realistic deployment benchmark would mix in 10k+ Western tracks. Out of scope here.
- Shorter queries are first-N truncations, not random re-cuts — known methodological gap from the Chang 2021 protocol; documented but not fixed.
- Recipe v3 trades epoch parity for convergence — trained 30 ep vs baseline's 10 ep. The improvement is recipe + 3× longer training combined. Individual ablations not run.
- 3 seeds is the minimum for stable pooled-McNemar; 5+ seeds 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 them.
- Tani ablation cell N = 11 (Hindustani: 0; Carnatic only) — per-section HR@1 claims for Tani are underpowered.
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 (sharply scoped to repo code, see github.com/ipritamdash/afp-indian-classical) |
See LICENSE, LICENSE-CODE, and ATTRIBUTION.md in this repo for full text.
Citation
@misc{afp-indian-classical-2026,
author = {Pritam Kumar},
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 pre-registered training-recipe improvement to NAFP (Chang et al. 2021).},
}
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},
}
Changelog
- v0.6 (2026-05-14): Major release.
- Recipe v3 training-recipe improvement —
68 % miss-rate reduction across benchmark (71 % on main_1s); pre-registered Bonferroni-significant on the two hardest cells (3 seeds × 30 epochs × BSZ=320, pooled McNemar p = 3.18 × 10⁻⁶ on main_1s, p = 0.0046 on ablation_1s) - Added NMFP-ckpt-100 (Araz et al. ISMIR 2025) as system #5 — establishes HR@1 = 1.000 ceiling across all 8 cells
- Added pre-registered negative results: Intervention 2 (per-artist mean subtraction) and hubness post-processing (InvSoftmax + CSLS)
- All 4 query lengths now exposed per (system × split); v0.5 only had 10 s
- Added pooled-McNemar CSV + protocol markdowns for full transparency
- README rewritten for clarity
- Recipe v3 training-recipe improvement —
- v0.5 (2026-05-12): Post-mortem audit fixes.
has_twin_in_libraryrecomputed via work_mbids only (was works-text); Dejavuref_stopfixed;score.pyno_matchdenominator fixed; NAFPnp.argsortmade stable. Headline HR@1 unchanged; twin/no-twin breakdown shifted (165 → 130 confirmed twins). - v0.4 (2026-05-12): NAFP added as 4th system.
- v0.3 (2026-05-12): Refreshed ablation result parquets to match v0.2 manifest. Length-curve added.
- v0.2 (2026-05-12): F5 fix (NFKD ablation bucketing); 624 → 632 ablation queries.
- v0.1 (2026-05-12): Initial release with 3 classical systems on 624 ablation queries.
Attribution
See ATTRIBUTION.md for full credits to Saraga / NAFP / NMFP / FMA upstream authors.
- Downloads last month
- 76