Phonetic spellchecker

This program takes an English-y word and returns the dictionary word(s) with most similar pronunciation. It catches a subset of misspellings invisible to traditional spellcheckers — namely those that “look different, but sound the same”.

Usage

python spellcheck.py

Example run

Enter ‘q’ to quit.

Enter a word: akshulie
Did you mean: actually

Enter a word: kuzz-zyn
Did you mean: cousin

Enter a word: zookeeny
Did you mean: zucchini

Enter a word: q

How it works

The idea is simple: map lexical sequences (words) to phonemic sequences (pronunciations), then look up the nearest phonemic sequence in the dictionary.

We map words to pronunciations with a sequence-to-sequence transformer trained on the CMU Pronouncing Dictionary. Sequences are compared with Levenshtein distance.

Dependencies

The following packages need to be installed:

1. nltk — for the words corpus.
2. editdistance — fast C implementation of Levenshtein distance.
3. torch — neural network library. Version 1.6.0 or higher.
4. torchtext — used to process data. Note: the Example and Field classes will soon be deprecated.

Optimizations

There is lots of room for experimentation and improvement.

• Sequence translation is greedy — could be improved with beam search.
• Edit distance search is exhaustive and likely too slow for some applications (see https://norvig.com/spell-correct.html for alternatives).
• Almost no hyperparameter search conducted while training the model. Doubling dimensionality (from 128 to 256) yields about 2% accuracy improvement but at the cost of quadrupling size. The bundled seq2seq.pt file is ~5 MB and >20 MB seemed excessive.
• Vanilla transformer model — no fancy modifications.
• Syllabic information stripped away for simplicity — could be incorporated.
• Label smoothing may also improve accuracy.

For your own use

To incorporate this work into a larger system, simply instantiate a Recommender and call recommend(), which takes a word and returns a list of matching words.

Disclaimer: Few checks on malformed input. Also, .py files not organized as importable modules.

Acknowledgements

This small program would not be possible without the decades-long work that has gone into the CMU Pronouncing Dictionary.

I would also like to thank Ben Trevett and Aladdin Persson, whose transformer-from-scratch code I copied extensively.