DECtalk babbling ...
These sort-of sound like English words. I trained the smallest possible language model (a Markov chain of length 2) on the phonemes generated when DECtalk read a chapter of a book. The model picks the next sound as a random one likely to follow the previous ones.