The architecture of the aspect speech recognition system, a speaker-dependent connected-word recognizer is described. As an integral part the system contains an adaptable interface to application programs, based on partitioned finite-state semantic grammars and a message passing scheme to track the state transitions of the application. The interface design is demonstrated to be effective, both in reducing perplexity and in the context-dependent interpretation of recognized word sequences. Finally, a PC-based real-time implementation is presented, which consists of two separate processes: the connected-word recognizer and a window-oriented sample application written in an object-oriented LISP extension.