Mapping graphemes to phonemes in CMUDict

Question

I'm trying to make a fun little "Halloweenify" feature where a user types in his or her name and gets a scary version. (Julie becomes "Ghoulie", Robert becomes "Macabert").

I have a huge list of Halloween works and a huge list of names, all with their phonemes from CMUDict. (Let's put aside for a second uncommon names.) For example:

"julie": "JH UW1 L IY0"
"ghoul": "G UW1 L"

It's relatively easy to write some code to align the "UW1 L" in these words. But since I don't explicitly know that the "IY0" in "Julie" corresponds to "ie," I can't get from "Ghoul" to "Ghoulie."

I think what I need is a version of CMUDict, or any other robust grapheme-to-phoneme dictionary, that splits up both the orthography and the phonemes, like so:

"j u l ie": "JH UW1 L IY0"
"r o b er t": "R AA1 B ER0 T"

That way, it would be easy to reconstruct the original name after adulterating it to make it scary.

Has anyone seen such a dictionary? I'm also open to more generic g2p solutions, but since this runs in a browser I can use any live heavy-duty ML methods. Thanks!

Answer 1

This is a fairly hard problem (well, 2: the g2p problem, and the word-formation problem). Per-grapheme:phoneme (fun fact: such mappings are sometimes called ‘graphones') dictionaries for English don’t (can’t) exist, because there’s lots of ambiguity and issues of multiple n-gram sizes (as in your example, ‘ie’ = /IY/, but sometime ‘i’ = /IY/, or /IH/, or whatever). Dealing with stress placement is even worse, and possible not solvable by the method I suggest below.

I had some success with Phonetisaurus, a more generic g2p (and, if you reverse the inputs, p2g) tool. It needs to be trained on a dictionary like you describe (orthography list with matched phonemic list).

I guess you’ll also have to match the spooky versions for stress and rhyming, but I don’t know of an existing solution for that.

The most realistic solution is for humans to prepare these spooky versions ahead of time (by hand); the distribution of names is such that you could get reasonably good coverage for most people with a reasonable amount of work.

Answer 2

There's a paper by Jiampojamarn and Kondrak: Letter-Phoneme Alignment: An Exploration that comes with software: m2m-aligner. This can work on cmudict (but you need to reformat cmudict). I ran it without the emphasis markers and got:

j|u|l|i:e|      JH|UW|L|IY|
g|h|o:u|l|      G|_|UW|L|
r|o|b|e:r|t|    R|AA|B|ER|T|

Reproducibility

Here's what I ran:

git clone --depth 1 \
    https://github.com/letter-to-phoneme/m2m-aligner.git
cd m2m-aligner
perl -pi -e 's/-lgcc_s//' makefile
make
curl -O \
 http://svn.code.sf.net/p/cmusphinx/code/trunk/cmudict/cmudict-0.7b

Here's how I converted the cmudict format to what m2m-aligner wanted:

import re
f = open('cmudict.txt', 'w')
for line in open('cmudict-0.7b', 'r', encoding='latin1'):
    if not re.match(r'^[a-zA-Z\d\s]+$', line): continue
    word, phonemes = line.strip().split(' ', 1)
    if not re.match(r'^[a-zA-Z]+$', word): continue
    print(' '.join(word.lower()), '\t',
          re.sub(r'\d\b', '', phonemes.strip()),
          sep='', file=f)

And then I ran

./m2m-aligner --delX --maxX 2 --maxY 2 -i cmudict.txt

The output is in cmudict.txt.m-mAlign.2-2.delX.1-best.conYX.align.

Answer 3

We just finished a project for which we developed a Phonics Engine (see paper https://www.researchgate.net/publication/280147388_Building_a_Phonics_Engine_for_Automated_Text_Guidance) that does exactly that. We have a dictionary (BrEng) that produces mappings such as b:b,a:@,tch:tS - both the code and the final dictionary are available but, unfortunately, the mapping code is very rough and ready and required lots of manual tweaks. Here's the code on GitHub: https://github.com/mpakarlsson/ilearnrw/tree/master/src/ilearnrw/languagetools/english.

Edit: The key files are probably just embedded in the Java code somewhere, so I started a separate repo for the language source files here: github.com/techczech/phonicsengine.