Skip to content

aparrish/pincelate

Repository files navigation

Pincelate

By Allison Parrish

Pincelate is a machine learning model for spelling English words and sounding them out, plus a Python module that makes it super simple to do fun and useful things with the model.

A quick example:

>>> from pincelate import Pincelate
>>> pin = Pincelate()
>>> pin.soundout("pincelate")
['P', 'IH1', 'N', 'S', 'AH0', 'L', 'EY1', 'T']
>>> pin.spell(["HH", "EH", "L", "OW"])
'hello'

Please see the documentation for more information!

I also did a series of tutorials on Pincelate for PyCon 2020.

Installation

Machine learning moves fast and breaks things, including backwards compatibility with models like this. So installation is a bit tricky. You'll need to install particular versions of Tensorflow and Keras to get Pincelate to work. (For this reason, I highly recommend installing Pincelate in a virtual environment or conda environment.)

This should do the trick:

pip install tensorflow==1.15.0 keras==2.2.5 "h5py<3.0.0"

After you've done this, you can install Pincelate:

pip install pincelate

This will install the code and the pre-trained model.

Pincelate requires Python 3.6 or later. (It might work on other versions, but I haven't tested it.) As of this writing, Python 3.8 and later are not yet supported (because of incompatibilities in some of the required libraries).

Model card

Following the schema suggested in Mitchell et al..

Model details

Pincelate was developed and trained by Allison Parrish over the course of 2019. The current version of the model is 0.0.1. The model consists of a pair of sequence-to-sequence recurrent neural networks that predict phonetic features from orthography, and orthography from phonetic features.

The model and accompanying code are available under an MIT open source license. See LICENSE in this repository.

If you make use of Pincelate in your research, please cite this repository.

Send questions or comments to <aparrish@nyu.edu>.

Intended use

The model has several primary intended uses:

  • Guess phonetic pronunciations of English words based on their spelling (grapheme to phoneme translation)
  • Guess English spellings of arbitrary phonetic pronunciations (phoneme to grapheme translation)
  • Provide a vector representation of an English word's phonetics, based on its spelling
  • Facilitate "transformations" of phonetics and spelling through manipulation of the model's internal state; e.g., "tinting" phonetics by adjusting the logits of the grapheme-to-phoneme model, "blending" words by averaging vector representations recovered from RNN hidden states, etc.

The model is designed to facilitate ease of use and creative tinkering, not fidelity or accuracy. For this reason (and others, outlined below), use of the model in situations where mistakes in automatically generated phonetic transcriptions or spellings could lead to critical breakdowns in communication (like text-to-speech synthesis) is not recommended.

The envisioned users are creative coders interested in using natural language processing as part of creative writing projects (e.g., computer-generated poetry).

Factors and metrics

This section is incomplete! At present, the model training process reports simple cross-entropy loss and accuracy scores on a validation dataset. A few goals for future versions of code implementing the training process:

  • Test against a dataset entirely held out of the training process
  • Test against a dataset of neologisms and non-standard English spellings not present in the CMU pronouncing dictionary
  • Separately evaluate words in categories pertaining to their semantics, etymologies and sociolinguistic use patterns, especially those pertaining to groups vulnerable to prejudicial treatment

Evaluation data and training data

The model is trained and evaluated on data from the CMU Pronouncing Dictionary ("CMUdict"), a freely available computer-readable phonetic dictionary of English words. At training time, the data undergo a train/validation split, with membership and proportion determined at training time (using a user-selectable random seed).

Internally, the model converts CMUdict phones to "phonetic features"; those features are automatically assigned to each phone based on a scheme proposed by Kirschenbaum.

Ethical considerations and caveats

The spellings and pronunciations in CMUdict reflect a dialect of English named in the dataset's documentation as "North American English," an equivalent to "Standard American English." As the model is trained on this dataset, the model's outputs can be expected to reproduce the phonetics and spelling conventions of this variety of English. The model does not attempt to accurately model other accents of English, or to model conventional methods of spelling those accents.

Orthographic variation, and phonetic variation that surfaces as orthographic variation, is not value neutral. In particular, "eye dialect" (the deliberate use of nonstandard spelling to draw attention to a word's pronunciation) can be used to mock and disparage speakers with particular accents or speech disorders. The model itself is not trained on any "eye dialect" spellings, but can be made to produce spellings that resemble eye dialect through careful manipulation of the model's internal state and decoding process. Applications making use of this model should take care to limit this affordance.

About

Easy to use ML model for spelling and sounding out words

Resources

License

Stars

Watchers

Forks

Packages

No packages published