This model was developed for the Google WAXAL ASR Challenge on Zindi.

waxal-w2vbert-linsna-seed46 is a Lingala and Shona CTC model, and the most decorrelated member of the family. It is a fresh fine-tune of facebook/w2v-bert-2.0, a 600M parameter encoder, at seed 46, with its own raw character vocabulary built from unnormalised transcripts, capitals and the punctuation CER scores included.

It shares its corpus and every hyperparameter with the seed-43 arm. Only the seed differs. That is the point: a fresh initialisation is the cheapest way to obtain a model whose errors are unlike the rest of the family, and error diversity is what a character-level vote converts into accuracy.

Measured over the 892 clip evaluation set, comparing greedy transcripts clip by clip:

Agreement below 13 per cent with every existing arm, while scoring comparably on its own, is the profile a vote wants from a new member.

30,987 rows after the speaker-disjoint holdout and the validation rows are removed.

Disclosure. The WAXAL Phase 1 test split is used as training data here. Phase 2 permits this and the organisers confirmed it publicly. It is gold reference text released by the organisers, not pseudo-labels, and it is not the Phase 2 test set.

Settings:

Trained with add_adapter: true and no inherited vocabulary, so this checkpoint's CTC token order differs from the lineage rooted at waxal-w2vbert-linsnalug-raw and it is not weight-average compatible with it. It combines by voting, not by averaging. Target normalization is NFC and whitespace collapse only: no lowercasing, no punctuation stripping, because CER is computed on the raw string and WER does not strip punctuation. The feature cache was held in RAM rather than on disk, which changes nothing about the result. Hardware: 1x NVIDIA RTX 5090 (32 GB VRAM), 32 CPU cores, 92 GB RAM.

Augmentation is applied on the fly in the data collator, so each epoch sees a fresh draw of the same clip:

The holdout is a speaker-disjoint carve of the WAXAL training data. It is not the competition test set, and it shares recording conditions and language variety with the training corpus, so it flatters any arm tuned on that corpus and gives no independent read on generalisation. Lower is better; the competition score is 1 - 0.5*(WER + CER).

No solo public-leaderboard score was recorded for this checkpoint.

Read the holdout figure with care. It is the best in the family, yet substituting this arm for the weakest member of a 26-member ensemble moved that ensemble from 0.766580 to 0.766477, which is slightly worse and well inside measurement noise. A member's solo quality is a poor predictor of what it contributes to a vote, and this checkpoint is a clean example of the gap.

Intended use is automatic speech recognition for Lingala and Shona photo-description speech: spontaneous descriptions of images, on consumer recording equipment, averaging 20.2 seconds per clip. It has not seen Luganda; for Luganda, start from waxal-w2vbert-linsnalug-raw, the only published checkpoint in this family that did.

This checkpoint is one component of an ensemble. It contributes to two of the seven ensembles that the scored submission, recipe p2n_mbr (public 0.766791, private 0.772552), selects across, anchoring one of them, and it is published so that result can be rebuilt in full. On its own it is an ordinary member of the family rather than the strongest one.

When transcribing a batch rather than a single clip, pass the attention mask through to the model. Clips in a batch have different lengths and the shorter ones are zero-padded; without the mask, self-attention treats that padding as real audio and the encoder output changes for every frame, not just the padded tail. Omitting it altered roughly half of all transcripts in a side-by-side check.