The best solution is out of scope for this website. Just apply an algorithm that converts binary data to human-readable pronounceable text.
A simple solution, much like base-64 or diceware, would be to download a dictionary, split binary ciphertext into chunk, and replace each chunk by selecting the nth word in the dictionary. Join the words together using spaces as separators. Punctuation can be added randomly.
The process would be reversed by removing punctuation, splitting on spaces, looking up the index of each word, then joining the binary format of the indices together.
Format preserving encryption is undesirable in this case. English prose is not a simple data format. It would be hard to convert one message to another without leaking some information about the content and/or format of the text.
Chances are also pretty good that readers wanting to implement the same kind of thing as you would roll their own format-preserving encryption algorithm which is little more than a substitution cipher. (Replacing plaintext words with pseudorandom dictionary words, for example.)
Substitution ciphers can be broken with paper and pencil. They put puzzles in newspapers next to Sudoku and comics where people are challenged to do exactly that.
The primary reason I suggest looking at reversible methods of encoding binary data as English text is because I think it's best to work to work on the process of encrypting and of formatting data as two totally independent problems.
When you use normal strong encryption (don't forget authentication), the ciphertext will be indistinguishable from a randomly generated bit sequence. It won't matter if the binary-to-text conversion reveals something about the input because the input is ciphertext. Even side channels in the conversion algorithm wouldn't be a problem if the encryption implementation isn't vulnerable to side-channel attacks.
Therefore, I think you would have better luck if you took that half of the question to StackOverflow.
You should not have high hopes for the results. Procedural generation of text is a difficult problem. It's of interest to AI researchers, spammers, news media, anyone with customer support chat, and artists. Even the most successful algorithms produce text that seems unnatural
It's really hard to generate text which can pass for something written by a real person. Luckily (or not, if you use a spam filter) it's also somewhat hard for computers to tell the difference between human-generated and computer-generated text.
As for what kind of suggestion I would give if you asked on StackOverflow...
I would take word frequency and n-gram frequency into account if the purpose of the algorithm was steganography. "To" is more likely to appear in a normal sentence than "meteorite" is. "Go to" is more likely than "jump bread".
Think of using the auto-complete function on a phone. You could write a random sentence that might look legitimate at first glance pretty easily. You could randomly select the first one or two letters of each word and select from the options given to you. An n-gram based data encoding algorithm might produce pretty similar results.
If steganography is not required (say the purpose of making ciphertext human readable was so that you could send a message over an audio channel or less-than-reliable text channel) then the simple word list encoding method is what I would try. I would make sure the word list doesn't contain homophones or even words with similar or confusing spellings.
(See one of EFF's short diceware word lists. There is one where the first three letters of a word uniquely determines which number it corresponds to and no two words have spellings which are too close to make spellcheck hard. Something like this would be more resilient to typos and misunderstandings.)