A model called Fish could identify both musical tracks and alternate versions of those tracks in benchmark tests, according to an arXiv preprint. The authors report that none of seven existing models met both accuracy and robustness requirements on the two tasks, while Fish showed unified operation with 10-second track-identification queries.
The Fish baseline was built from groups, or cliques, of 20-second audio segments containing different versions of the same musical phrase. The benchmark tested whether version identification - matching different versions of the same musical phrase - could also cover track identification while retaining accuracy and robustness when queries were manipulated or degraded.
A benchmark built around altered audio
The benchmark used Discogs-VI for both tasks, SHS100K for version identification and NMFP for track identification. It tested clean queries, manipulated queries, and queries that were both manipulated and degraded.
The databases contained about 116,000 Discogs-VI full tracks, 8,000 SHS100K full tracks and 95,000 NMFP full tracks. Evaluation covered about 372,000 full-track version-identification queries and 633,000 segment track-identification queries, more than one million queries in all.
Among the comparison systems were NMFP for track identification, CLEWS for version identification and CLAP, a language-audio representation model.
Clean and altered audio split the results
On clean Discogs-VI track tests, NMFP put the correct track first more often than Fish: 97.4% ± 0.1 versus 83.7% ± 0.2. Among the first 10 results, the figures were 99.7% ± 0.0 and 93.1% ± 0.1. After manipulation, Fish led on first-result matches, 74.4% ± 0.3 to 48.8% ± 0.3, and after manipulation plus degradation, 53.6% ± 0.3 to 47.2% ± 0.3. The paper reports the ± values as 95% confidence intervals.
For version identification, Fish posted the strongest overall results in the reported comparisons. Its mean average precision (MAP), an overall ranking score, was 0.687 ± 0.002 on clean Discogs-VI queries and 0.612 ± 0.002 after manipulation and degradation; on SHS100K, the corresponding scores were 0.806 ± 0.007 and 0.748 ± 0.007. Fish's normalized average rank (NAR) on Discogs-VI was lower than CLEWS's in all three conditions, including 9.1 ± 0.1 versus 11.3 ± 0.1 after both changes.
CLAP performed competitively on clean and manipulated track-identification tests but did not provide comparable version-identification performance, indicating that generic audio representations did not necessarily encode work identity.
The retrieval tests exposed a trade-off
In a separate retrieval-mechanics test, Fish's scores were 99.9% for tracks and 100.0% for versions when 10-second database segments were aligned with 10-second track-identification queries. With a 2.5-second boundary mismatch, the scores were 92.7% and 99.6%. The authors attribute the original track-identification gap to retrieval constraints rather than a limitation of the embedding space.
The reported scores showed a trade-off by segment duration. With track-identification queries fixed at 10 seconds, 10-second segments had first-result scores of 92.7% on clean queries and 74.6% after manipulation and degradation; 20-second segments had 91.9% and 69.2%. For version identification, MAP was higher with 20-second segments - 0.780 and 0.713 - than with 10-second segments, at 0.529 and 0.390, for the same two conditions.
What the benchmark leaves open
Those findings are bounded by the tested query settings. Track-identification evaluation used 10-second queries, while version-identification evaluation used full-track queries; performance for shorter clips remains open.
The authors state that they are sharing code, trained model weights and metadata, along with timestamps at 1-second resolution for 97 million pairs of 20-second audio segments. The material is grouped into 644,000 segment cliques from 317,000 tracks across 74,000 track cliques.
Taken together, the benchmark supports the feasibility of unified operation in the tested 10-second track-identification setting, but leaves open which context duration best balances track and version identification.
Paper data and sources
Original title: Unified Music Identification for Tracks and Versions
Authors: R. Oguz Araz, Joan Serrà, Yuki Mitsufuji et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text