In an industrial context, ad-hoc/manual testing strategies, using natural language, still seem to be highly prevalent since natural language descriptions are, more likely, easier to understand. Still, the lack of rigor can generate inaccurate tests. Aligned with other modern approaches, we promote the use of natural language descriptions with rigorously defined underlying semantics. As a distinguished feature of our approach, we cover the entire (direct engineering) testing process, from requirements to manual or automated test cases generated automatically. Requirements written in a controlled natural language are parsed, and their semantics are automatically modeled using the CSP process algebra. To address soundness and deal with different abstraction levels, we formalize the concept of a domain model, in which additional information, such as hierarchical composition and a dependence relation among test steps, is defined. Then, sound test cases are generated from the inferred scenarios using the cspio conformance relation. These test cases, still expressed in CSP, can then be linearized back to natural language to allow manual execution or directly translated into test scripts for automated execution.