If I wanted to produce plaintext like ciphertexts, I'd:
1. Encrypt the plaintext with a standard cipher, producing uniformly random bytes
2. Build a language model which predicts the probability of the next symbol (nowadays you'd probably use a neural network for that).
3. Run an arithmetic decoder (or similar) for that model on the ciphertext. This translates random bytes into language. The quality of the output depends on the quality of the language model.
1. Encrypt the plaintext with a standard cipher, producing uniformly random bytes
2. Build a language model which predicts the probability of the next symbol (nowadays you'd probably use a neural network for that).
3. Run an arithmetic decoder (or similar) for that model on the ciphertext. This translates random bytes into language. The quality of the output depends on the quality of the language model.