arabic-stt private data: manifests and language models
PRIVATE. Never make this repository public, and never copy any of its files into a public repository. The manifests hold the reference transcripts of research-licensed (SADA, MGB-2 / MGB-3 / MGB-5, NADI-2026, ArzEn, Mixat, GigaSpeech) and ND-licensed (Casablanca) corpora and of the leaderboard test sets; the language models are built from those transcripts.
| path | contents |
|---|---|
data/manifests/ |
lhotse CutSet manifests and their .meta.json (corpus splits, tiers, dev / test sets, the unlabeled pool, ta marbuta sidecars, reports) |
data/lm/ |
token-level KenLM binaries with their cards and lmplz logs; data/lm/text/: the LM text and its leakage reports |
The audio is not here: rebuild it from the pinned revisions (docs/data.md of the code repository).
Restore with uv run python scripts/sync_offbox.py restore --what data (docs/env.md, Off-box persistence);
the manifests store absolute audio paths, so run scripts/relocate_data.py when the checkout moved.
- Downloads last month
- 11