Apppublic

Nina-HK/luxlip-corpus

sourceHugging Facecc-by-nc-4.0updated 3mo agoView on Hugging Face

0likes

App README

LuxLip Corpus: A Luxembourgish Talking-Face Dataset for Visual Speech Recognition

Predicting Luxembourgish speech from silent talking-face videos.

Interactive demo for the LuxLip pilot corpus — browse real Luxembourgish talking-face clips, inspect the preprocessing pipeline (face crop → 88×88 grayscale model input), and read gold transcripts. Live AV-HuBERT transcription can be enabled by adding a fine-tuned checkpoint.

Access: sign-in required (username/password). See `ADMIN.md` for setup.

Prepared for SLT.

What you can do

—Examples — browse clips from the corpus (speaker, session, split, transcript)
—Try it — upload a short face video and see the model input preprocessing
—About — dataset stats and preprocessing pipeline details

Local development

bash

cd demo
cp .env.example .env.local   # set LUXLIP_DEMO_USER / LUXLIP_DEMO_PASSWORD
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python ../demo/prepare_examples.py          # local paths from Data/
python ../demo/prepare_examples.py --bundle  # bundle media for Spaces
./run_local.sh

Deploy to Hugging Face Spaces

bash

hf auth login
cd demo
LUXLIP_DEMO_USER=admin LUXLIP_DEMO_PASSWORD='choose-a-strong-password' ./deploy_hf.sh

Or set HF_SPACE=your-username/luxlip-corpus. Full admin instructions: `ADMIN.md`.