Constitutive descriptions of the orthotropic passive myocardium span a wide range of phenomenological models. However, subjective model choice often limits accuracy across multiple deformation modes and microstructural organization. We propose a data-driven framework that autonomously discovers both the form of the strain energy function and fiber dispersion in ventricular walls. The approach employs an incompressible orthotropic constitutive neural network that maps eight strain invariants to a scalar-valued strain energy function, which enables systematic exploration of billions of admissible material models [1,2]. We train the network using experimental data from uniaxial extension as well as simple shear tests [3]. The learning procedure minimizes errors between modeled and measured stress components, while lasso regularization promotes sparsity and objective model selection. Training across all available deformation modes ensures consistent identification of material behavior and avoids bias toward individual loading conditions. The discovered models accurately reproduce the experimental responses and reveal a compact constitutive model dominated by a single isotropic invariant and anisotropic invariants associated with preferred myocardial directions. The framework further identifies effective dispersion parameters directly from mechanical data and captures microstructural variability. We further examine the model robustness with respect to data noise, initialization, and microstructural assumptions. Our results show that constitutive neural networks provide an objective tool for material model and dispersion discovery in the myocardium. The proposed methodology supports reliable cardiac tissue modeling, improves predictive capability under complex loading states, and advances data-driven constitutive modeling for computational biomechanics and cardiac simulation.