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

waxal-w2vbert-linsna-afrivoicemix continues the waxal-w2vbert-linsnalug-raw checkpoint, a fine-tune of facebook/w2v-bert-2.0, on a single mixed corpus: WAXAL Lingala and Shona gold together with both AfriVoice corpora, present in every batch for the whole run. It is the deliberate counter-experiment to the two-stage specialists, which see exactly the same external audio but sequenced, AfriVoice first and WAXAL gold last. Because the ordering is the only intended difference, this arm is what turned a plausible curriculum heuristic into a measured result, and the result was negative for mixing: 0.2905 on the holdout for mixing throughout, against 0.2785 for the sequenced Lingala specialist and 0.1616 for the sequenced Shona specialist. It is kept in the submission's core ensembles as a decorrelated member, not as a strong solo model. It also inherits the parent checkpoint's raw character vocabulary unchanged, capitals and the punctuation CER scores included.

All three sources are mixed into one training set and there is no stage boundary. Luganda is dropped from this arm because the corrected Phase 2 test set was measured to contain none. Rows and hours above are per-source totals from the corpus build; the speaker-disjoint holdout and the validation rows are removed before training, and the duration and characters-per-second filters drop outliers.

Settings:

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

Targets are raw for the same reason as every other arm here: CER is computed on the raw string and WER is computed on lowercased but otherwise unmodified text, so punctuation errors are charged twice and a model trained on stripped targets is capped on both metrics.

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 speakers' 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).

This repository holds a retrained checkpoint, not the weights used in the submission. The original was lost to a disk cleanup, so the arm was retrained from the same config at the same seed. It is a genuinely different model: it disagrees with the original on 766 of the 892 test clips, 85.9 per cent. The 26-member ensemble built with it nonetheless changed on only 50 of 892 rows and scored 0.764774 against 0.764915 for the original, a difference of 0.000141 against a standard error of about 0.0013. The scored submission, recipe p2n_mbr, is touched even less: rebuilt from published weights it matches the submitted file on 883 of 892 rows, because five of its seven candidate ensembles carry this arm at half weight and the per-clip selection can fall back on the other candidates.

Read the two holdout figures with care. The retrained checkpoint's 0.2599 looks far better than the original's 0.2905, yet the ensemble containing it scored fractionally worse on the leaderboard. The two holdout numbers may not share a carve, and in any case a member's solo quality is a poor predictor of what it contributes to a vote.

This is a negative result and is reported as one. The intended benefit of mixing was that WAXAL text would anchor the output conventions while AfriVoice supplied in-domain acoustics, avoiding the punctuation drift the sequenced specialists show at the end of their AfriVoice stage, which their refresh epoch then has to undo. That benefit did not appear in the score: the sequenced specialists finished ahead on both languages. The rule taken from the comparison, external acoustics first and in-style gold last, was then applied to every later arm, including the Whisper one. The refresh is not free money either: on an arm supplemented with already-punctuated FLEURS text it changed nothing (0.2809 before, 0.2812 after), which is consistent with transcription convention rather than acoustics being the thing the refresh repairs.

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 is not a general-purpose model for either language and has not been evaluated outside this domain. The competition targets Lingala, Shona and Luganda; the corrected Phase 2 test set was measured to contain only Lingala and Shona, and Luganda is dropped from this arm's corpus, so it should not be given Luganda audio.

By solo holdout error this is the weakest of the w2v-BERT arms in this solution, and that is the point of it. It exists as a controlled comparison first and as an ensemble member second: its training distribution differs from every other arm, so its errors are correspondingly decorrelated, and source diversity was measured to be worth more than parameter diversity (four members from four checkpoints gained 0.0040, six members from two checkpoints gained 0.0006). The submission's 26-member character-level ROVER core scored 0.764915, and the scored final submission, recipe p2n_mbr, selects per clip across seven such ensembles and scored 0.766791 public and 0.772552 private. Solo strength is not the only criterion for membership: the blank-penalty re-decodes over-generate words and lose score in isolation (0.743 against 0.746 for the same checkpoint) while contributing +0.0104 in total, because a character vote can filter a spurious word and can never recover a missing one. The limit of that argument was measured as well: genuinely weak checkpoints did not help, and a 21-member vote including them scored 0.764152 against 0.764759 for a 25-member vote of strong sources.

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.