Devrim, Mehmet OsmanKirisoglu, Serdar2025-10-112025-10-1120250003-682X1872-910Xhttps://doi.org/10.1016/j.apacoust.2025.110887https://hdl.handle.net/20.500.12684/22019Gold purity determination is a critical part of quality control in jewellery and industrial production. Traditional methods (chemical analyses, X-ray fluorescence, etc.) can be costly, time consuming and destructive. This study aims to present an audio-based non-destructive alternative for gold purity classification. Mel-Frequency Cepstral Coefficients (MFCC) features delta-delta extracted from audio signals were analysed with 10 different machine learning algorithms (Support Vector Machines (SVM), Decision Trees, Logistic Regression, Random Forest, KNearest Neighbour (KNN), Naive Bayes, Gradient Boosting, AdaBoost, XGBoost, LightGBM). The dataset was divided into training, test and validation subsets; features were normalised with StandardScaler and the generalization performance of the models was optimised with 5-fold cross-validation. In the comparison of performance metrics (accuracy, precision, recall, F1-score), it was observed that SVM (94.58%) and Logistic Regression (93.75%) models were superior to other algorithms, especially in capturing subtle differences between classes. Confusion matrices detail the success of the model in discriminating 14, 22 and 24 carat classes. This study proves that the use of audio data in gold purity analysis has the potential for a fast, repeatable and non-destructive solution in industrial applications.en10.1016/j.apacoust.2025.110887info:eu-repo/semantics/closedAccessGold purity gradesNon-destructive testingAudio classificationMachine learningArticle2402-s2.0-105010297615WOS:001554458300001Q1Q1