w2vbert-lin-waxal-aug

Compliant Lingala W2V-BERT 2.0 CTC model for the Google WAXAL ASR Challenge. Fine-tuned from facebook/w2v-bert-2.0 on WAXAL train split only, with case + punctuation kept in the vocab, a speaker-disjoint validation split, and noise + speed augmentation for out-of-domain (Phase-2) robustness.

Best validation (speaker-disjoint, honest)

Validation speakers are disjoint from training speakers, so this predicts fresh-audio (Phase-2) performance rather than the inflated numbers a leaking split gives.

• Combined (0.5·WER + 0.5·CER): 0.3098
• WER: 0.4394 | CER: 0.1802
• step 5600 (epoch 6.79)

Training results

epoch	step	val loss	WER	CER	combined
0.48	400	2.8141	0.9864	0.7944	0.8904
0.97	800	1.2500	0.4904	0.1949	0.3427
1.45	1200	1.0934	0.4898	0.2035	0.3466
1.94	1600	1.1093	0.4587	0.1850	0.3218
2.42	2000	1.1206	0.4685	0.1888	0.3287
2.91	2400	1.0381	0.4457	0.1821	0.3139
3.39	2800	0.8541	0.4555	0.1829	0.3192
3.88	3200	1.0334	0.4425	0.1812	0.3119
4.36	3600	1.0769	0.4386	0.1811	0.3099
4.85	4000	1.0821	0.4406	0.1806	0.3106
5.33	4400	1.0916	0.4452	0.1822	0.3137
5.82	4800	1.0029	0.4417	0.1801	0.3109
6.30	5200	1.0550	0.4449	0.1813	0.3131
6.79	5600	0.9753	0.4394	0.1802	0.3098
7.27	6000	1.1131	0.4437	0.1810	0.3124

Data & compliance

• Training data: WAXAL Lingala train split only. The Phase-1 test split is never read (no test audio or transcriptions used at any point).
• Validation: a speaker-disjoint slice carved from the labelled data (no speaker appears in both train and validation).
• Base model: facebook/w2v-bert-2.0 (openly available).
• Augmentation: Gaussian noise at randomised SNR + speed perturbation.
• Labels: case and punctuation preserved (minimal NFC + whitespace cleanup only).