Voice characteristics are an emerging, non-invasive biomarker for heart failure. This study develops a machine learning pipeline to differentiate patients with suspected heart failure from those with a confirmed diagnosis using vocal features alone, drawing on 240 patients (50 suspected, 190 confirmed) from six European medical centres who completed a multi-test voice-recording protocol. From 490 extracted voice features, collinearity filtering and LASSO regularization reduced the set to 22 key biomarkers; combined with SMOTE class-balancing, an Extra Trees classifier achieved 78.4% accuracy and a macro-F1 score of 0.76, with 89.5% sensitivity for confirmed HF, supporting refined vocal biomarkers as a reliable HF screening tool. This work was presented at the 5th Serbian International Conference on Applied Artificial Intelligence (SICAAI 2026), Kragujevac, Serbia, and was carried out within the STRATIFYHF project.