Unified Music Identification for Tracks and Versions

R. Oguz Araz; Joan Serrà; Yuki Mitsufuji; Xavier Serra; Dmitry Bogdanov

cs.SDAug 20, 2026

Unified Music Identification for Tracks and Versions

Authors: R. Oguz Araz, Joan Serrà, Yuki Mitsufuji, Xavier Serra, Dmitry Bogdanov

Organizations: Music Technology Group, Universitat Pompeu Fabra, Barcelona · Sony AI · Sony Group Corporation

Abstract

Given a music database, track identification (TI) retrieves the exact track matching an audio excerpt, whereas version identification (VI) retrieves its musical versions. Traditionally, the two tasks have been addressed separately. However, as every track is its own closest version, we investigate whether VI can subsume TI. This requires VI systems to be robust to both signal manipulation and audio degradation. We therefore propose a unified benchmark that evaluates accuracy and robustness on each task. Comparing seven existing models on this benchmark, we show that none of them are both accurate and robust on both tasks. We then train a baseline model targeting both tasks and show that a unified system is possible with 10 s TI queries. Lastly, we characterize the two retrieval constraints that limit our model's TI performance. We envision extending this unification to other music identification tasks.

Figures & tables

Figure 1: Music editing-playback-capture chain. Signals highlighted in red are used as queries during evaluation.

Figure 2: Pairwise segment-level distance matrices of 3 musical versions, computed on audio segments of 20 s with 1 s hop, obtained with CLEWS. Similar regions located by our method are outlined by the white contour lines.

Metric	Model	NMFP Test						Discogs-VI Test
Metric	Model	Clean		Manip.		Manip. & Deg.		Clean		Manip.		Manip. & Deg.
hit rate@1 ( $\uparrow$ )	GraFPrint	83.7 $\pm$	0.2	41.3 $\pm$	0.3	11.7 $\pm$	0.2	71.0 $\pm$	0.3	36.3 $\pm$	0.3	12.0 $\pm$	0.2
	PeakNetFP	94.8 $\pm$	0.1	31.2 $\pm$	0.3	2.4 $\pm$	0.1	95.9 $\pm$	0.1	32.1 $\pm$	0.3	2.7 $\pm$	0.1
	NMFP	96.9 $\pm$	0.1	48.7 $\pm$	0.3	45.5 $\pm$	0.3	97.4 $\pm$	0.1	48.8 $\pm$	0.3	47.2 $\pm$	0.3
	CLAP	89.9 $\pm$	0.2	43.2 $\pm$	0.3	3.2 $\pm$	0.1	84.5 $\pm$	0.2	32.1 $\pm$	0.3	1.9 $\pm$	0.1
	CoverHunter	10.8 $\pm$	0.2	5.0 $\pm$	0.1	1.5 $\pm$	0.1	10.8 $\pm$	0.2	6.0 $\pm$	0.1	1.5 $\pm$	0.1

Table 1: track identification benchmark results. The

\pm

symbol denotes a 95% confidence interval.

Metric	Model	Discogs-VI Test						SHS100K Test
Metric	Model	Clean		Manip.		Manip. & Deg.		Clean		Manip.		Manip. & Deg.
MAP@10 k ( $\uparrow$ )	GraFPrint	0.014 $\pm$	0.001	0.007 $\pm$	0.000	0.005 $\pm$	0.000	0.010 $\pm$	0.001	0.007 $\pm$	0.000	0.007 $\pm$	0.000
	NMFP	0.010 $\pm$	0.000	0.005 $\pm$	0.000	0.009 $\pm$	0.000	0.010 $\pm$	0.001	0.006 $\pm$	0.000	0.008 $\pm$	0.000
	CLAP	0.019 $\pm$	0.001	0.010 $\pm$	0.000	0.002 $\pm$	0.000	0.027 $\pm$	0.001	0.021 $\pm$	0.001	0.009 $\pm$	0.001
	CoverHunter	0.031 $\pm$	0.001	0.019 $\pm$	0.001	0.006 $\pm$	0.000	0.129 $\pm$	0.006	0.086 $\pm$	0.005	0.036 $\pm$	0.003
	ByteCover2	0.273 $\pm$	0.002	0.208 $\pm$	0.002	0.164 $\pm$	0.002	0.469 $\pm$	0.008	0.357 $\pm$	0.008	0.276 $\pm$	0.007

Table 2: version identification benchmark results. The

\pm

symbol denotes a 95% confidence interval.

Model	Misalignment (s)	Track hit rate@1	Version hit rate@1
Fish	$0.00$	95.0	98.5
Fish	$2.50$	91.9	98.2
NMFP	$0.00$	99.7	99.8
NMFP	$0.25$	96.7	99.8

Table 3: Effect of boundary misalignment between query and database segments on track identification performance. Database segments match each model’s training context duration; information mismatch is present for Fish. Misalignment values are upper bounds.

Misalignment (s)	Track hit rate@1	Version hit rate@1
0.0	99.9	100.0
2.5	92.7	99.6

Table 4: Effect of boundary misalignment on track identification performance for Fish with 10 s database segments, which eliminates the information mismatch. Misalignment values are upper bounds.

Dur. (s)	track identification		version identification
Dur. (s)	Clean	Manip. & Deg.	Clean	Manip. & Deg.
10	92.7	74.6	0.529	0.390
20	91.9	69.2	0.780	0.713

Table 5: Effect of segment duration on the track identification and version identification performance of Fish, applied to the database tracks and to version identification queries; track identification queries are 10 s in both settings. Track hit rate@1 and MAP@10 k are reported for track identification and version identification, respectively.

Data	N.M.	track identification		version identification
Data	N.M.	Clean	Manip. & Deg.	Clean	Manip. & Deg.
$\mathcal{D}_{1}$	SH	$86.5$	$40.9$	$0.551$	$0.359$
$\mathcal{D}_{2}$	SH	$86.1$	$45.7$	$0.611$	$0.469$
$\mathcal{D}_{3}$	SH	$81.3$	$46.7$	$0.697$	$0.552$
$\mathcal{D}_{4}$	SH	$84.5$	$51.6$	$0.686$	$0.550$
$\mathcal{D}_{4}$	H	90.3	64.6	0.738	0.645

Table 6: Effect of training data construction on Fish. hit rate@1 and MAP@10 k metrics are reported. N.M.: Negative mining strategy; SH: semi-hard, H: hard.

Music Information Retrieval

Music Information Retrieval develops computational methods for finding, classifying, describing, and analyzing music from audio, scores, and metadata.

See all 153 papers

Turning music identification into a neural forward pass
Jun 15, 2026Muhammad Taimoor Haseeb, Ahmad Hammoudeh, Gus XiaMusic Information Retrieval
Music X Lab, Mohamed Bin Zayed University of Artificial Intelligence, UAE
Towards Robust Version Identification in the Wild: A Dataset, Benchmark, and Fine-Tuning Study
Aug 5, 2026Simon Hachmeier, R. Oguz Araz, Dmitry Bogdanov +2Music Information Retrieval Cross-Dataset Generalization
Berlin School of Library and Information Science, Humboldt-Universität zu Berlin, Germany · Music Technology Group, Universitat Pompeu Fabra, Barcelona, Spain
Training Music Sample Identification Models on Real Sample Pairs
Sep 18, 2026R. Oguz Araz, Joan Serrà, Xavier Lizarraga-Seijas +3Audio Representation Learning Music Information Retrieval
Music Technology Group, Universitat Pompeu Fabra, Spain · Sony Europe · Sony CTC America

Jun 15, 2026J/K move · Enter open · S save

Turning music identification into a neural forward pass

Muhammad Taimoor Haseeb, Ahmad Hammoudeh, Gus Xia

Music X Lab, Mohamed Bin Zayed University of Artificial Intelligence, UAE

Music Information Retrieval

Open paper AskPDF arXiv

Search, a foundational operation in computer science, maps a query to a matching item in a collection. It is typically implemented as a System-2 like, rule-based pipeline in which a key is computed, an index is probed, and candidates are verified. By contrast, human recognition resembles a System-1 like, associative model of identity recovery, in which even partial cues can trigger a recall without explicitly enumerating, ranking, or even accessing discrete candidates. Here, we show that music sound identification, a difficult search problem, can be performed in a single neural feed-forward pass by a generative transformer. Trained on an audio dataset, the model predicts the corresponding track identifier from a short audio excerpt. This approach surpasses state-of-the-art acoustic fingerprinting, with the largest gains for short audio segments (1 second), demonstrating the method is not only viable but advantageous. Moreover, it reduces external storage to 0.33% of the baseline footprint and improves inference latency by 2.3x (p95). Furthermore, the model can reject queries for unseen tracks, supporting open-set operation while reducing misattribution risk. Using music track identification as an example, this work reframes search, bringing it closer in spirit to human associative recognition and away from algorithmic database lookup.

Abstract

Music Information Retrieval

Explore similar work

Turning music identification into a neural forward pass

Towards Robust Version Identification in the Wild: A Dataset, Benchmark, and Fine-Tuning Study

Training Music Sample Identification Models on Real Sample Pairs