One way of thinking about speech production is using a spreading activation architecture. Say I want to produce the word "rooster." I have a meaning to be expressed (feathery avian animal loud...etc). These semantic representations send activation down to various matching words (bird, rooster, chicken...etc). As these words collect evidence for themselves, they send activation down to their corresponding sounds (bird sends activation to the sounds /b/, /r/, and /d/; rooster sends activation to /r/, /u/, /s/, /t/ ..etc). Once any node (word, sound..etc) receives enough evidence for itself, it is selected. There are nuances about how and when nodes become selected and how activation is passed, but this is the gist. In addition, there is noise in the system based on words that were previously said, upcoming words, words that are semantically related and so on. This noise introduces the possibility for speech errors. Some errors are more likely than others depending on the amount of evidence they have collected. For example, in the example I presented here, the /r/ sound got more activation than the other sounds because it is collecting activation because it appears once in the word 'bird' and twice in the word 'rooster.' This sound might be more or less likely to be produced erroneously compared to sounds like /s/. These error probabilities explain why we are more likely to make real word errors, and why tongue twisters are so tricky. See the following article for more information.
Dell, G. S. (1986). A spreading-activation theory of retrieval in sentence production. Psychological review, 93(3), 283.