ABSTRACT Varying moisture content in wood lead to problems in the automatic selection of optimised machining parameters. Therefore, production in the wood sector can be improved by the automatic identification of moisture classes during machining. In this work, wood milling was monitored by a novel optical microphone to detect differences in wood moisture classes. A full factorial randomised design was executed along with four moisture classes (dried, conditioned, wet and frozen) and two cutting speeds as independent variables. The measured signals were segmented by extracting information from each cut and were transformed to the frequency domain. A supervised univariate feature ranking based on χ ²-tests was applied to select the most important frequencies. Machine Learning was applied to the selected spectral data to classify the cutting speeds and moisture classes. After testing several classification models, ensemble models with regression-tree learners were the best-performing model type achieved an average validation accuracy of 97.2%. The model explanation was done by Local Interpretable Model-Agnostic Explanation based on regression trees as a simple model, where the frequencies that are responsible for the separation of the data could be identified and were used for retraining the model, achieving a validation accuracy of 90.5%.