New algorithm identifies songs from degraded audio with 98.4% accuracy
Researchers cracked song ID from noisy audio using a novel algorithm hitting 98.4% accuracy
A new optimization framework called Construct-Merge-Solve-Adapt (CMSA) achieves state-of-the-art results on the Longest Filled Common Subsequence problem, a challenging NP-hard task with applications in bioinformatics and, as a novel engineering contribution, song identification from degraded audio excerpts. The adaptive method iteratively builds promising subproblems, solves them with an external black-box solver, and refines them using feedback from previous rounds. In experiments on standard and new large-scale benchmarks, it outperformed four known approaches and matched the proven optimum on 1,486 of 1,510 instances with known optimal solutions—98.4% of those cases.
- Algorithm achieves 98.4% accuracy on 1,510 test instances, matching optimal solutions for 1,486 cases
- CMSA framework uses iterative construction and external solvers to handle large-scale problems
- New benchmark dataset with larger instances exposes limitations in existing evaluation methods
Why It Matters
Enables reliable song identification from noisy recordings, transforming music recognition and copyright enforcement