We present a method to determinize and minimize speech recognition networks where any relevant sub-network can be switched on/off at runtime without recompilation. We show how the general algorithm for determinizing general weighted transducers can be applied for this task.