Differentiable digital signal processing has recently given rise to novel data-driven approaches to artificial reverberation. Within this framework, a delay-network model, such as a Feedback Delay Network (FDN), can be optimized to match the time-frequency sound energy decay of a target acoustic environment. Typically, such optimization relies on the availability of measured Room Impulse Responses (RIRs). However, obtaining accurate RIR measurements is not always feasible in practical scenarios. In this work, we show that it is possible to optimize the FDN parameters directly on musical recordings, without requiring dedicated impulse response measurement sessions. Under controlled experimental conditions, the proposed method achieves performance comparable to RIR-based optimization, making it a promising alternative for applications such as audio production and virtual acoustics.