In the modern fast-changing digital world, the rapid spread of misinformation through rumors is a serious issue, causing anxiety, harming reputations, and increasing societal polarization. Identifying rumors is inherently difficult due to their evolving nature, making traditional methods less effective. To address this challenge, we conducted a comparative analysis of various machine learning models and natural language processing techniques for multi-source rumor detection using datasets from PHEME, Snopes, and others. Pre-processing involved POS tagging (Unigram, Bigram, Trigram, HMM, and SpaCy), lemmatization (Morphological, Affix, and Radix Tree), and feature extraction. We evaluated ten machine learning models, including Logistic Regression, SVM, Random Forest, KNN, and ensemble techniques, using accuracy, precision, recall, and F1-score. Among all models, SVM consistently outperformed others with an accuracy of 92.3% and an F1-score of 91.8%, demonstrating its robustness in rumor classification. Random Forest and Logistic Regression also showed competitive performance with accuracy scores of 89.7% and 88.5%, respectively. Meanwhile, Naive Bayes and KNN exhibited lower and inconsistent results, with accuracy dropping to 79.2% and 76.8%, depending on the feature extraction methods. Despite being theoretically strong, models like GBM and AdaBoost struggled due to a notable precision-recall trade-off, achieving F1-scores of 85.6% and 83.9%, respectively.