Liver cancer, including hepatocellular carcinoma (HCC), cholangiocellular carcinoma (CCC), and metastases, presents diagnostic challenges during surgery due to its infiltrative nature. Accurate intraoperative classification and margin assessment are crucial for improving outcomes. Current methods, like frozen section analysis, are time-consuming and subjective, necessitating rapid, objective alternatives. This study assessed fiber-based attenuated total reflection infrared (ATR IR) spectroscopy combined with supervised machine learning for intraoperative liver tumor classification based on a holistic biochemical signature approach. Fresh liver tissue from 69 surgical patients was analyzed using a probe consisting of Ge ATR crystal and silver halide fibers. Supervised algorithms reliably classified normal tissue and tumor subtypes (HCC, CCC, metastases) using cross-validation and independent test sets. Normal liver tissue was distinguished primarily by differences in glycogen content and structural compactness of tumor tissue. Normal and tumor tissues were differentiated with a sensitivity of 0.89 and a specificity of 0.92. The accuracy of spectroscopic classification is 0.90. The three-group classification of tumor subtypes also yielded an average accuracy of 0.90. HCC is characterized by a higher glycogen content compared to CCC and metastases and can be identified spectroscopically with high reliability. CCC showed distinct protein-associated spectral signatures, while metastases exhibited unique profiles reflecting their different origins. In a minority of cases, misclassifications occurred, indicating potential for further refinement. Fiber-based ATR IR spectroscopy in combination with machine learning provides a rapid, objective, and highly accurate intraoperative tool for liver tumor classification. This label-free biochemical approach may enhance surgical precision and reduce recurrence risks across the full range of solid tumor entities.