As speech translation systems become increasingly integrated into real-world applications, there is a growing need for comprehensive evaluation frameworks that reflect both linguistic accuracy and functional performance. Existing evaluation methods often overlook acoustic difficulties and error propagation which significantly affect system input and output. This study proposes a unified evaluation framework tailored for both cascading and end-to-end speech translation systems. It considers acoustic difficulties factors and error propagation and allows for reward mechanisms in cases where translation stages correct upstream errors. A case study evaluating a cascading system using varied-accent English speeches demonstrates that the proposed framework effectively captures system performance across diverse conditions. While the framework is only on system level and dependent on subjective human ratings, it provides a solid foundation for future development of objective, fine-grained evaluation standards.