A critical step in aircraft design is to specify the vibration levels that onboard equipment must be able to withstand throughout their life for correct operation. These levels are essentially determined experimentally, as predicting the dynamic behavior of a complete aircraft requires computational means and methods that are currently difficult to access. We present a data-driven methodology that leverages flight test accelerometer data to produce a predictive model. This model, based on artificial neural networks, performs a multi-output multivariate regression to estimate vibration spectra from a set of aircraft general parameters. While vibration environment results from several excitation sources, this methods does not require any source characterization. In previous work (Février et al., 2023, submitted to AIAA Journal) we developed a predictive tool for level flight phases. The method is now extended to all flight phases, which include complex maneuvers. This significantly increases the dimension of the problem, as unsteady flight dynamics are characterized by a large number of time-dependent parameters. Different network architectures, including recurrent neural networks, were tested in order to better handle temporal dynamic behavior. A comparison with baseline models is performed over multiple test protocols, from standard train-test random split to flight enveloppe extrapolation. We show that our model is more suited to real industrial needs, as it performs better on domain-specific tasks. An interpretation of different predictions is provided from the study of the most influential parameters using the SHAP method. It suggests that our model is able to identify the key features that generate vibration levels. The proposed methodology has several applications in the aircraft industry such as anomaly detection and flight test optimization, providing new tools to address such predictive challenges.