Submission for the IWSLT 2026 Low Resource Track
# IWSLT 2026 Low-Resource Speech Translation Submission
Dual-pipeline system for the IWSLT 2026 Low-Resource ST shared task, targeting all 10 language pairs with a unified approach that combines cascaded and end-to-end speech translation with MBR ensemble decoding.
## Architecture
```
┌──────────────────────────────┐
│ Source Audio │
└──────────┬───────────────────┘
│
┌───────────────┴───────────────┐
▼ ▼
┌─────────────────────┐ ┌─────────────────────────┐
│ Pipeline A: Cascade │ │ Pipeline B: End-to-End │
│ │ │ │
│ Whisper large-v3 │ │ SeamlessM4T v2 large │
│ (LoRA, MTL ASR+ST) │ │ (LoRA fine-tuned) │
│ │ │ │ │
│ ▼ │ │ Direct ST output │
│ NLLB-200 1.3B │ │ + N-best hypotheses │
│ (intra-distill.) │ │ │
│ │ │ └────────────┬────────────┘
│ N-best hyps │ │
└─────────┬───────────┘ │
│ │
└────────────┬───────────────────┘
▼
┌─────────────────────────┐
│ MBR Decoding (chrF) │
│ 50+50 hypotheses │
│ Select best per utt. │
└────────────┬────────────┘
▼
Final Translation
```
## Language Pairs
| Pair | Direction | Data | Pipeline Notes |
|------|-----------|------|----------------|
| arn-spa | Mapuzugun → Spanish | ~130h | E2E primary (not in NLLB) |
| bem-eng | Bemba → English | ~180h | Cascade + E2E |
| bho-hin | Bhojpuri → Hindi | ~26h | Pseudo-translation trick |
| ckb-eng | C. Kurdish → English | ~30h | Cascade + E2E |
| gle-eng | Irish → English | ~25h | Cascade + E2E |
| ibo-eng | Igbo → English | ~20h | Cascade + E2E |
| hau-eng | Hausa → English | ~20h | Cascade + E2E |
| que-spa | Quechua → Spanish | ~58h | Cascade + E2E |
| ca-en | Catalan → English | ~30h | Whisper direct ST available …