Classification of CD- and LP-Derived Audio Using Hand-Crafted Acoustic Features and Interpretable Machine Learning

Sandhya Rani K. M.
School of Computational Sciences, Swami Ramanand Teerth Marathwada University (SRTM University), Nanded, Maharashtra, India
Email: sandhya.kiran.rao@gmail.com (S. R. K. M.)
S. D. Khamitkar
Email: s.khamitkar@gmail.com (S. D. K.)
Parag U. Balachandra
Email: srtmun.parag@gmail.com (P. U. B.)
ORCID: Sandhya Rani K. M. 0009-0009-2027-8513
Parag U. Balachandra 0000-0001-5422-9423
*Author for correspondence: srtmun.parag@gmail.com

Abstract

Compact discs (CDs) and long-playing records (LPs) reproduce audio through different mechanisms, creating measurable differences in digitised music signals. We classified CD- and
LP-derived audio using hand-crafted acoustic features and interpretable machine learning models. A private dataset comprising 17,800 non-overlapping 10-s segments from 597 songs
was evaluated using Random Forest, XGBoost, support vector machine (SVM) and k-nearestneighbour (k-NN) classifiers. Because segments from the same song are correlated, we used song-grouped cross-validation, independent validation and a held-out test set to prevent data leakage. Random Forest and XGBoost achieved 99.96% segment-level accuracy and correctly
classified all 90 held-out songs. Feature importance, SHAP and ablation analyses indicated that spectral-harmonic and psychoacoustic characteristics provided the strongest discrimination. An auxiliary format-conversion analysis showed that the pre-trained XGBoost model retained the LP label for 94.13% of LP segments converted from WAV to MP3 and the CD label for 93.90% of CD segments decoded from MP3 to WAV. These findings demonstrate a strong separation between the evaluated CD- and LP-derived audio classes. However, the classes contained different songs and were represented in different formats, so musical content, encoding, mastering and recording-chain differences may have influenced the results. Stricter content-matched and codec-controlled studies are needed to isolate the effects associated with the physical medium.

Keywords: audio provenance, compact disc, vinyl record, acoustic features, Random Forest, XGBoost, SHAP, machine learning