Skip to content
<awesome-gh-repos>
vic

aparrish/phonetic-similarity-vectors

Source code to accompany my paper "Poetic sound similarity vectors using phonetic features"
Source code at GitHub https://github.com/aparrish/phonetic-similarity-vectors
aparrish/phonetic-similarity-vectors.json
{
  "createdAt": "2017-07-31T22:05:11Z",
  "defaultBranch": "master",
  "description": "Source code to accompany my paper \"Poetic sound similarity vectors using phonetic features\"",
  "fullName": "aparrish/phonetic-similarity-vectors",
  "homepage": "",
  "language": "Jupyter Notebook",
  "name": "phonetic-similarity-vectors",
  "pushedAt": "2017-10-25T20:56:01Z",
  "stargazersCount": 171,
  "topics": [],
  "updatedAt": "2025-04-13T01:59:33Z",
  "url": "https://github.com/aparrish/phonetic-similarity-vectors"
}

Poetic sound similarity vectors using phonetic features

Section titled “Poetic sound similarity vectors using phonetic features”

This is the source code used to implement the algorithms, experiments and applications in my paper, “Poetic sound similarity vectors using phonetic features.” The source code is written in Python.

You can read the paper here. EXAG posted a summary of the paper as well, and you can watch this talk that I gave at Strange Loop 2017 that discusses much of the paper’s content in an accessible way.

Dependencies

Section titled “Dependencies”

Almost everything should be covered in a standard Anaconda install, i.e.:

  • pandas
  • matplotlib
  • numpy
  • scikit-learn

You’ll also need spaCy and Annoy for some of the applications, which you can install with pip or through conda-forge. For spaCy:

conda config --add channels conda-forge
conda install spacy
python -m spacy download en_core_web_md

For Annoy:

conda config --add channels conda-forge
conda install python-annoy

(Note: You don’t need any of this stuff if you just want to play around with the pre-computed vectors! Though I definitely recommend using a fast nearest-neighbor search library like Annoy.)

Files

Section titled “Files”

CMU Dict and pre-computed vectors

Section titled “CMU Dict and pre-computed vectors”

The file cmudict-0.7b-with-vitz-nonce contains the most current (as of this writing) version of the CMU Pronouncing Dictionary, edited to include the “nonce” words from Vitz and Winkler (1973).

The file cmudict-0.7b-simvecs contains pre-computed vectors for all of the words in cmudict-0.7b-with-vitz-nonce. This is probably the file you want if you just want to play around with the vectors! The vectors are formatted just like the CMU Pronouncing Dictionary (i.e., word in all caps, two spaces, then space-separated values per dimension).

Python source

Section titled “Python source”

Run generate.py to create your own set of vectors from a CMU Pronouncing Dictionary source file. Pass the dictionary as input and redirect the output to your desired file, e.g.:

PYTHONIOENCODING=latin1 python generate.py <cmudict-0.7b-with-vitz-nonce >cmudict-0.7b-simvecs

Note that you must specify latin1 encoding when running this script (unless your dictionary uses some other character set).

The similarity.py script is a quick script for checking your vectors. Pass the file with your similarity vectors as a command line argument, and the program will respond to every line of standard input with the most similar items from the embedding:

$ python similarity.py cmudict-0.7b-simvecs
loading...
done.
ELEPHANT
['ELEPHANT', "ELEPHANT'S", 'ELEPHANTS', "ELEPHANTS'", 'ENTOFFEN', 'UFFELMAN', 'UNRUFFLED', 'MUFFLE', "ENTOFFEN'S", 'KALAFUT']
AARDVARK
['AARDVARK', 'AARDVARKS', 'AARGH', 'ARGH', 'ARC', 'ARK', 'ARB', 'ARTCARVED', 'ALSGAARD', 'ARCHARD']
BADGER
['BADGER', 'BADER', 'BATHER(1)', 'BADGERED', 'BISER', "BADGER'S", 'BADGERS', 'BADDERS', 'MADAR', 'MADDER']
DOLPHIN
['DOLPHIN', 'DOLPHINS', "DOLPHINS'", 'DALFEN', 'GALVEN', 'DONELSON', 'GALVAN', 'JARVIS', 'GALVESTON', 'DARTH']

Hit ^D when you’re done.

Finally, featurephone.py contains a few helper functions that help build the bigram analysis of the dictionary.

Jupyter notebooks

Section titled “Jupyter notebooks”
  • experiment.ipynb contains the code to run the experiments comparing the phonetic similarity reported by the embeddings in my paper to the similarity scores obtained from human subjects in Vitz and Winkler (1973).
  • some-applications.ipynb contains a number of playful and poetic experimental applications of the phonetic similarity embeddings (including sound analogies, averages, symbolism tinting, etc.)

Still forthcoming: An example of how to use the embeddings for longer stretches of text, as in the “random walk” example in the paper.

License

Section titled “License”

See LICENSE, which applies to everything in this repository except the copy of the CMU Pronouncing Dictionary, which is used under the terms of their license (included in the header of the file), and the data from Vitz and Winkler, for which I claim fair use.