Gloss-Based Pipeline for Spoken to Signed Language Translation

a text-to-gloss-to-pose-to-video pipeline for spoken to signed language translation.

• Demos available for:

  • 🇩🇪 Swiss German Sign Language 🇨🇭
  • 🇫🇷 French Sign Language of Switzerland🇨🇭
  • 🇮🇹 Italian Sign Language of Switzerland 🇨🇭

• Paper available on arxiv, presented at AT4SSL 2023.

Install

pip install spoken-to-signed

Then, to download a lexicon, run:

download_lexicon \
  --name <signsuisse> \
  --directory <path_to_directory>

Usage

For language codes, we use the IANA Language Subtag Registry. Our pipeline provides multiple scripts.

To quickly demo it using a dummy lexicon, either open it in Colab:

or run it locally with some installation steps first:

git clone https://github.com/ZurichNLP/spoken-to-signed-translation
cd spoken-to-signed-translation

pip install .

text_to_gloss_to_pose \
  --text "Kleine Kinder essen Pizza in Zürich." \
  --glosser "simple" \
  --lexicon "assets/dummy_lexicon" \
  --spoken-language "de" \
  --signed-language "sgg" \
  --pose "quick_test.pose"

Text-to-Gloss Translation

This script translates input text into gloss notation.

text_to_gloss \
  --text <input_text> \
  --glosser <simple|spacylemma|rules|nmt> \
  --spoken-language <de|fr|it> \
  --signed-language <sgg|ssr|slf>

Text-to-Gloss-to-Pose Translation

This script translates input text into gloss notation, then converts the glosses into a pose file.

text_to_gloss_to_pose \
  --text <input_text> \
  --glosser <simple|spacylemma|rules|nmt> \
  --lexicon <path_to_directory> \
  --spoken-language <de|fr|it> \
  --signed-language <sgg|ssr|slf> \
  --pose <output_pose_file_path>.pose

Words missing from the lexicon are fingerspelled letter-by-letter by default; pass --disable-fingerspelling to skip them instead.

Add --coverage-info to print how each token was matched, color-coded in the terminal (green: lexicon, yellow: language backup, orange: fingerspelling, red: unmatched), or --coverage-stats <file.json> to save the same per-token information as JSON.

Text-to-Gloss-to-Pose-to-Video Translation

This script translates input text into gloss notation, converts the glosses into a pose file, and then transforms the pose file into a video.

Note: Video generation requires the pose-to-video package with pix2pix and upscaler:

pip install 'pose-to-video[pix2pix,simple_upscaler] @ git+https://github.com/sign-language-processing/pose-to-video'

text_to_gloss_to_pose_to_video \
  --text <input_text> \
  --glosser <simple|spacylemma|rules|nmt> \
  --lexicon <path_to_directory> \
  --spoken-language <de|fr|it> \
  --signed-language <sgg|ssr|slf> \
  --video <output_video_file_path>.mp4

HTTP service

pip install '.[server]'
MODEL_VERSION=local hypercorn spoken_to_signed.server:app --bind 0.0.0.0:8080

POST /tokens-to-gloss accepts atomic words/spans with POS and optional morphology:

{
  "spoken_language": "en",
  "signed_language": "ase",
  "tokens": [
    {"word": "What", "gloss": "what", "pos": "PRON"},
    {"word": "is", "gloss": "be", "pos": "AUX"},
    {"word": "your", "gloss": "your", "pos": "PRON"},
    {"word": "name", "gloss": "name", "pos": "NOUN"},
    {"word": "?", "gloss": "?", "pos": "PUNCT"}
  ]
}

Returns sentences containing the original objects in your name what ? order, plus indexes: [[2, 3, 0, 4]]. All annotations pass through unchanged; no merging in the caller. glosser is rules (default, English → ASL) or simple (identity). morphology is a list of spaCy feature dictionaries, one per source token in a span. These are mechanical rules, not fluent ASL. Unknown words remain for downstream fingerspelling; punctuation remains for sentence boundaries. TODO: batch API.

For model-based reordering, install .[server,gpt], set OPENAI_API_KEY at runtime, and send "glosser": "gpt". This uses gpt-5.6-luna with reasoning disabled (override model with OPENAI_MODEL); only rules-approved omissions are allowed. Rules/simple need no API key.

POST /senses-to-gloss accepts senses (the WSD document with tokens, synsets, and entities) plus the same language/glosser fields. It groups multiword spans before glossing and returns the same response. Indexes refer to grouped candidates; each candidate carries its original start_token/end_token, exact-span senses and entities, and source annotations for later lookup/fallback. Unknown words survive. Overlapping spans prefer the widest meaning (earlier on ties); constituent senses are never treated as senses of the whole phrase. No dictionary lookup happens here.

POST /gloss-to-pose accepts already ordered tokens and the same language fields, returning binary application/pose. It uses the existing lookup and concatenation, excluding pos: "PUNCT". Optional: fingerspelling=true, anonymize=false, source (PostgreSQL video-ID prefix). Configure LEXICON_PATH for CSV or DATABASE_URL for PostgreSQL (takes precedence); without either, pose requests return 503. For PostgreSQL, install .[server,postgres,gcs] and use a read-only role. The optional SQLPoseLookup backend migrates the old models captions-table lookup and reads gs://sign-mt-poses; it is not dictionary-api synset lookup. Glosser/health requests need no database. See /docs for the full API schema.

docker build --build-arg MODEL_VERSION=local -t spoken-to-signed .
docker run --rm -p 8080:8080 spoken-to-signed

Releases publish ghcr.io/sign-language-processing/spoken-to-signed-translation:<tag>. The image includes PostgreSQL/GCS/GPT dependencies. PORT defaults to 8080; Hypercorn supports HTTP/1.1 and HTTP/2 (h2c). Configure HTTP/2 upstream in the gateway too. /health returns version; successful API responses include X-Model-Tag, set by MODEL_VERSION (baked into release images), with a suffix identifying the configured GPT model and base URL. Deploy internally; auth and caching belong to the gateway. Caller routing and the old caption-maintenance job still need migrating before retiring the function in models.

Methodology

The pipeline consists of three main components:

1. Text-to-Gloss Translation

   Transforms the input (spoken language) text into a sequence of glosses.

• Simple lemmatizer,
• Spacy lemmatizer: more accurate, but slower lemmatization, covering fewer languages than simple,
• Rule-based word reordering and dropping component and
• Neural machine translation system.

2. Gloss-to-Pose Conversion

• Lookup: Uses a lexicon of signed languages to convert the sequence of glosses into a sequence of poses.
• Pose Concatenation: The poses are then cropped, concatenated, and smoothed, creating a pose representation for the input sentence.

3. Pose-to-Video Generation

   Transforms the processed pose video back into a synthesized video using an image translation model.

Supported Languages

Online Playgrounds

We have two available:

• sign.mt is a web interface of a translation system.
• research.sign.mt is an overview of sign language processing literature.

Citation

If you find this work useful, please cite our paper:

@inproceedings{moryossef2023baseline,
  title={An Open-Source Gloss-Based Baseline for Spoken to Signed Language Translation},
  author={Moryossef, Amit and M{\"u}ller, Mathias and G{\"o}hring, Anne and Jiang, Zifan and Goldberg, Yoav and Ebling, Sarah},
  booktitle={2nd International Workshop on Automatic Translation for Signed and Spoken Languages (AT4SSL)},
  year={2023},
  month={June},
  url={https://github.com/ZurichNLP/spoken-to-signed-translation},
  note={Available at: \url{https://arxiv.org/abs/2305.17714}}
}