A Lightweight Selective Classification Framework for Music-Centric YouTube Videos Using Audio Repetition and Metadata
Year: 2026
Type: report
Abstract: Music-centric video selection for karaoke differs from detecting the presence of music: films, trailers, cookingvideos, and travel montages can contain highly repetitive soundtracks while serving a non-musical purpose.This exploratory study documents a lightweight pipeline combining traditional chroma-based repetitioncoverage, short-lag periodicity, YouTube category exclusions, and a reject option for human review. Noneural network, large language model, or pretrained audio classifier is used for prediction. Audio indicatorsare normalized using selection-set percentiles, combined with fixed candidate weights, and evaluated withselection-calibrated dual thresholds. The research progressed from a 50-item pilot experiment to threeexploratory batches totaling 323 videos, with a 273-item speech-reduced replay and a 229-item historicalholdout. On that repeatedly inspected holdout, the pure-coverage configuration with category filtering andselective Entertainment review automatically classified 190 videos, rejected 39, and made 11 automatic errors:coverage 82.97% and selective risk 5.79%. Entertainment protection intercepted eight music-backed falsepositives while additionally rejecting one correctly classified song. These results do not meet the strict risktarget below 5%. Repeated reuse of historical data, provisional labels, unresolved recording-independenceevidence, and convenience sampling prevent independent validation claims. The contribution is an auditableengineering combination and empirical diagnosis, not a new theory of chroma, repetition, or selectiveclassification.
Author: Wang, Suyuan
Institution: RWTH Aachen University
Language: English
Topic: Music and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225904303
