SignWriting

Python utilities for SignWriting.

Installation

pip install git+https://github.com/sign-language-processing/signwriting

Or with Docker:

docker build --platform linux/amd64 --tag signwriting:python .
docker run --platform linux/amd64 --rm -p 9090:8080 -e PORT=8080 signwriting:python

Utilities

signwriting.formats

This module provides utilities for converting between different formats of SignWriting. We include a few examples:

1. To parse an FSW string into a Sign object, representing the sign as a dictionary:

from signwriting.formats.fsw_to_sign import fsw_to_sign

fsw_to_sign("M123x456S1f720487x492")
# {'box': {'symbol': 'M', 'position': (123, 456)}, 'symbols': [{'symbol': 'S1f720', 'position': (487, 492)}]}

2. To convert a SignWriting string in SWU format to FSW format:

from signwriting.formats.swu_to_fsw import swu2fsw

swu2fsw('𝠃𝤟𝤩񋛩𝣵𝤐񀀒𝤇𝣤񋚥𝤐𝤆񀀚𝣮𝣭')
# M525x535S2e748483x510S10011501x466S2e704510x500S10019476x475

signwriting.tokenizer

This module provides utilities for tokenizing SignWriting strings for use in NLP tasks¹. We include a few usage non-exhaustive examples:

1. To tokenize a SignWriting string into a list of tokens:

from signwriting.tokenizer import SignWritingTokenizer

tokenizer = SignWritingTokenizer()

fsw = 'M123x456S1f720487x492S1f720487x492'
tokens = list(tokenizer.text_to_tokens(fsw, box_position=True))
# ['M', 'p123', 'p456', 'S1f7', 'c2', 'r0', 'p487', 'p492', 'S1f7', 'c2', 'r0', 'p487', 'p492'])

2. To convert a list of tokens back to a SignWriting string:

tokenizer.tokens_to_text(tokens)
# M123x456S1f720487x492S1f720487x492

3. For machine learning purposes, we can convert the tokens to a list of integers:

tokenizer.tokenize(fsw, bos=False, eos=False)
# [6, 932, 932, 255, 678, 660, 919, 924, 255, 678, 660, 919, 924]

4. Or to remove 'A' information, and separate signs by spaces, we can use:

from signwriting.tokenizer import normalize_signwriting

normalize_signwriting(fsw)

signwriting.visualizer

This module is used to visualize SignWriting strings as images. Unlike sutton-signwriting/font-db which it is based on, this module does not support custom styling. Benchmarks show that this module is ~5000x faster than the original implementation.

from signwriting.visualizer.visualize import signwriting_to_image

fsw = "AS10011S10019S2e704S2e748M525x535S2e748483x510S10011501x466S20544510x500S10019476x475"
signwriting_to_image(fsw)

To use the visualizer with the server, you can hit: https://signwriting-sxie2r74ua-uc.a.run.app/visualizer?fsw=M525x535S2e748483x510S10011501x466S2e704510x500S10019476x475

signwriting.utils

This module includes general utilities that were not covered in the other modules.

1. join_signs joins a list of signs into a single sign. This is useful for example for fingerspelling words out of individual character signs.

from signwriting.utils.join_signs import join_signs_vertical

char_a = 'M507x507S1f720487x492'
char_b = 'M507x507S14720493x485'
result_sign = join_signs_vertical(char_a, char_b)
# M510x518S1f720490x481S14720496x496

2. canonicalize rewrites a sign's symbols in a canonical order (by category, then top-to-bottom, left-to-right), centers the sign, and tightens its box, without changing the rendered image. See utils/canonicalize.

from signwriting.utils.canonicalize import canonicalize

canonicalize('M521x547S20310506x500S33100482x443')
# M521x554S33100482x482S20310506x539

signwriting.fingerspelling

This module is used to generate spelling data from a list of characters.

from signwriting.fingerspelling.fingerspelling import spell

word = "Hello"  # any string of characters
language = "en-us-ase-asl"  # long language code, as defined in the fingerspelling README
spell(word, language)
# M515x563S11502477x437S14a20492x457S1dc20484x477S1dc20484x512S17620492x547

To use the fingerspelling with the server, you can hit: https://signwriting-sxie2r74ua-uc.a.run.app/fingerspelling?text=hello&signed_language=ase

signwriting.mouthing

This module is used to generate SpeechWriting from spoken words.

from signwriting.mouthing.mouthing import mouth

word = "Hello"  # any string of characters, preferably valid words
language = "eng-Latn"  # supported languages under "Language Support" at https://pypi.org/project/epitran/
mouth(word, language)
# M557x518S34700443x482S35c00469x482S34400495x482S34d00521x482

Note: Installing English support for epitran requires extra steps, see "Install flite" at mouthing/README.md.

To use the mouthing with the server, you can hit: https://signwriting-sxie2r74ua-uc.a.run.app/mouthing?text=hello&spoken_language=eng-Latn

Cite

@misc{moryossef2024-signwriting, 
    title={Utilities for SignWriting},
    author={Moryossef, Amit},
    howpublished={\url{https://github.com/sign-language-processing/signwriting}},
    year={2024}
}

References

2023. SignBank+: Preparing a Multilingual Sign Language Dataset for Machine Translation Using Large Language Models.