Heart failure (HF) remains a global health challenge [1], yet traditional diagnostic procedures are often limited by high costs and the need for specialized equipment. To improve accessibility, focus has shifted toward non-invasive biomarkers, specifically vocal characteristics, which can manifest detectable alterations in systemic pathologies [2]. This study focuses on an extraction of novel voice biomarkers and its utilization in development and training of a neural network for the diagnosis of HF. Voice biomarkers are extracted from audio recordings of patients who performed five different tests: spontaneous speech, reading of a text, numbers counting up, numbers counting down and maximum phonation test. These audio files were then processed and main vocal characteristics were extracted. In total 94 vocal characteristics were obtained for each task, including jitter, shimmer, loudness, cepstral and glottal features, etc. In order to mitigate small size and imbalance of the dataset, a couple of techniques were used: custom focal loss function, generating synthetic samples using ADASYN and L2 regularization. These structured voice characteristics are then used as input data for a novel multi-branch 1D convolutional neural network. This model treats features from each vocal task independently before fusing them for a final prediction. Each of the 5 parallel branches (one for each task) uses convolutional filters to find patterns in the feature space and at the end creates a concentrated representation of the features for that specific task. At the end, the final output is a fusion that merges the vectors from all 5 branches into a combined feature vector that is used for final classification by a traditional dense neural network. This model shows strong predicting capabilities, achieving accuracy of 84.62% on the test dataset. This demonstrates that automated speech analysis can serve as a highly accurate, non-invasive screening tool for heart failure.