WAXAL Checkpoints (private backup)
PRIVATE checkpoint backup for the Google WAXAL ASR Challenge — pushed to free DICO disk quota; pull
subfolders on demand. Each subfolder is a Wav2Vec2ForCTC fine-tune (char vocab) on WAXAL train (lin/sna/lug,
joint, speaker-disjoint holdout).
| subfolder | base | note |
|---|---|---|
w2vbert_multi |
facebook/w2v-bert-2.0 | CHAMPION (best clean+robust): band 0.1775 / corrupt 0.2250 |
xlsr_multi |
facebook/wav2vec2-xls-r-300m | XLS-R baseline (0.1862 / 0.2456) |
xlsr_aug |
facebook/wav2vec2-xls-r-300m | + MCT augmentation (robustness: degradation +0.0484) |
xlsr_robust |
facebook/wav2vec2-xls-r-300m | + full MCT |
xlsr_ext |
facebook/wav2vec2-xls-r-300m | + FLEURS external |
xlsr_pseudo |
facebook/wav2vec2-xls-r-300m | + 3-model pseudo-labels (student) |
Scoring metric = 0.5·WER + 0.5·CER on the speaker-disjoint holdout / MUSAN+RIR corrupted band.
Pull one
from huggingface_hub import snapshot_download
snapshot_download("anyantudre/waxal-checkpoints", allow_patterns="w2vbert_multi/*", local_dir="ckpt")
Bases are Apache-2.0; trained on google/WaxalNLP (respect its license). No Phase-1 test data used.