Shakibyzn/transformer-dgs-phoenix14t

Transformer DGS Phoenix14t

A minimal 3-layer Transformer for Sign Language Translation (SLT), trained on PHOENIX-2014T. It translates video features (like S3D) into German text using the google-bert/bert-base-german-cased tokenizer.

Intended Use & Performance

This model is intended for educational purposes. It demonstrates how to map continuous sign language embeddings to natural language sentences.

Baseline Performance:

• ~13 BLEU using standard S3D features.
• ~18 BLEU using features from neccam/slt.

Quick Start

Note: This model uses custom code. You must use trust_remote_code=True.

from transformers import AutoModel, AutoTokenizer

# Load Model & Tokenizer
model = AutoModel.from_pretrained("Shakibyzn/transformer-dgs-phoenix14t", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("Shakibyzn/transformer-dgs-phoenix14t", trust_remote_code=True)

Train and Inference

train.py and inference.py scripts are provided in the repository to demonstrate the train/test loop. This can be easily adapted for other sign language datasets or feature extractors (e.g., I3D/CLIP). It handles the loading of features and labels, and the standard training/validation loop.