anyantudre/waxal-pseudo

WAXAL Pseudo-Labels (3-model agreement)

PRIVATE working artifact for the Google WAXAL ASR Challenge — not for redistribution.

High-confidence pseudo-labels for the WAXAL unlabeled split, produced by a 3-model agreement cascade: a clip is kept only when the fine-tuned champion (w2v-BERT-2.0 CTC) and XLS-R-300m agree (CER ≤ 0.12), and the fine-tuned Omnilingual-ASR-300M independently confirms the champion transcript (CER ≤ 0.22). omni is architecturally diverse (different corpus/tokenizer), so it decorrelates the same-family champion↔XLS-R agreement. The label is the champion's transcript.

config (lang)	pseudo-labels
lin (Lingala)	see split
sna (Shona)	see split
lug (Luganda)	see split

Fields

id, audio (16 kHz), text (champion transcript), agree_xlsr (CER champ↔XLS-R), agree_omni (CER champ↔omni), dur (seconds).

Provenance & license

Audio is derived from google/WaxalNLP (unlabeled split). This repo re-hosts it privately only as a working backup; obtain the clean labeled data from google/WaxalNLP and respect its license. Contains no Phase-1 test transcriptions.

Load

from datasets import load_dataset
ds = load_dataset("anyantudre/waxal-pseudo", "lug", split="train")