SignCorpus — Nepali Sign Language (pilot release)
Multimodal NSL clips produced by the SignCorpus pipeline (Ingestion → Multimodal Extraction → Labeling & QC → Release), synchronizing RGB video, MediaPipe Holistic pose, audio, OCR-recovered captions, ASR transcript, pseudo-gloss, and NLLB-200 translations — filtered by the Quadruplet Contrastive Correspondence (QCC) metric.
- Run ID:
run_20260819_203918_bf2262 - Source video duration: 151.2s
- Clips accepted: 1 / 2 detected
- Unique signers: 1
- Pseudo-gloss vocabulary: 67 tokens
- OCR / ASR coverage: 100% / 100%
Models used
| Component | Checkpoint |
|---|---|
| ASR | milanakdj/whisper-medium-nepali-final |
| OCR | EasyOCR (['ne', 'en']) |
| Translation | facebook/nllb-200-distilled-600M |
| Text embedding | sentence-transformers/LaBSE |
| RGB embedding | openai/clip-vit-base-patch32 |
| Pose | MediaPipe Holistic |
⚠️ Known limitations (pilot release)
- QCC weights are priors (manuscript Table 3), not yet calibrated on human-reviewed data.
- Single-video pilot — the statistics above describe this run, not a full corpus.
- Signer identities are anonymized (
signer_id), but facial appearance remains inherently identifying — see the manuscript's Ethical Considerations (§5.4) before redistributing.
Full run manifest: run_manifest.json
- Downloads last month
- 11