Skip to content

Pronounceable Word Language

Authors
Joe Starr, Ph.D. image

Joe Starr, Ph.D.

I specialize in computational knot theory. I’m also a professional embedded software engineer.

What?

Some research disciplines dealing with human perception use English words as a base dataset for study. Unfortunately, simply selecting English words from the OED can introduce bias in a study. In these cases bias can be controlled by ensuring that words in the base dataset are sampled from the language of pronounceable in English but non-dictionary words (PEND).The task of deriving words in the PEND language isn't straight forward. There are two issues that arise when you attempt this task. First, how do you ensure that the derived word is actually pronounceable? Second, how do you ensure the word is not a bonafide English dictionary word?

The first issue is probably best addressed by some grammar based on phonetics pronounce ability. Unfortunately, at this point phonetics are beyond my knowledge base. Instead, we use a less sophisticated approach taken from an answer on Stack Overflow12. This approach uses bigrams (two Latin letters) and trigrams (three Latin letters) as the basis for forming words in the target language. In our grammar a word always starts with a bigram, transitions in the grammar then convert that bigram into a trigram. The conversion happens by selecting biased random choice a third character. Naturally the bigrams and trigrams found in English are not equally likely. We will use this difference in frequency as an integer weight representing its frequency in the Google n-gram dataset3. If we continue to derive from this grammar we arrive at infinitely long "pronounceable" strings. To control the length of the word we control the size of the stack. A stack of size \(n\) will produce a word of length \(n+1\). This combined with a calling script that randomly selects a length from the Google word length data 3 again weighted by frequency.

Length Frequency
0 0
1 26
2 622
3 4615
4 6977
5 10541
6 13341
7 14392
8 13284
9 11079
10 8468
11 5769
12 3700
13 2272
14 1202
15 668
16 283
17 158
18 64
19 40
20 16
21 1
22 5
23 2

The second issue is much simpler to solve. Unfortunately, actually implementing the solution must be done in a postprocessing step by a separate tool. Once we have derived a word we simply check if it's in an (or many) dictionaries.

Special Care?

As described above special needs to be taken when configuring the stack size.

What Happens When the Stack Is Full?

When the stack is full an empty string is produced a string containing the current bigram.

File

The file containing this grammar is rather large but can be found on GitHub.

  1. https://stackoverflow.com/questions/25966526/how-can-i-generate-a-random-logical-word ↩

  2. Paul . Wordle. 2014. Accessed 2026-03-11. URL: https://github.com/pcrov/Wordle. ↩

  3. Patrick Juola. Google Books Ngrams. In Laurie A. Schintler and Connie L. McNeely, editors, Encyclopedia of Big Data, pages 517–521. Springer International Publishing, Cham, 2022. doi:10.1007/978-3-319-32010-6_359. ↩↩