How should we understand speech recognition models and large models? How are their roles divided? Why do we need to configure a large model even when we have a speech recognition model?