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

waxal-w2vbert-linsna-soup5 is not a trained checkpoint. It is the uniform weight average of five checkpoints from the waxal-w2vbert-linsnalug-raw lineage, produced by averaging their parameter tensors elementwise. No gradient step was taken to create it and no data was seen. It covers Lingala and Shona.

Averaging is only meaningful because all five members share one vocabulary and one CTC token order. Every member either is waxal-w2vbert-linsnalug-raw or continues from it with vocab_dir pointing back at it, so position k of the CTC head means the same character in all five. Averaging checkpoints whose vocabularies were built independently, for instance the seed-43 and seed-44 arms, would produce a meaningless model; those arms combine by voting instead.

The result is the strongest w2v-BERT arm apart from the distilled one, and it is the checkpoint that waxal-w2vbert-linsna-distilled continues from.

The two _p2 members are pseudo-label rounds over the WAXAL unlabeled pools, refreshed on WAXAL gold. They are published only as ingredients of this average, not as standalone repositories. Self-training on those unlabeled pools was flat in both languages when measured directly, and snaspec_p2 was one of the checkpoints whose blank-penalty members took the vote from 0.764476 down to 0.764152. Neither is strong enough to vote on its own account.

Averaging is the way to use a checkpoint that is not good enough to vote. A weight average commits to one set of parameters and therefore one transcript, so a mediocre ingredient is diluted; a vote gives every member a say in every character slot, so a mediocre member injects its errors directly.

No data was used to produce this checkpoint. Transitively, through its five members, it derives from:

The pseudo-labels are our own models' output over WAXAL's unlabeled pools. They are not test audio. Self-training on those pools was flat in both languages when measured directly, which is why the _p2 checkpoints survive only as soup ingredients.

Uniform average, weight 0.2 each, over every floating-point parameter tensor. Buffers and the processor are taken from waxal-w2vbert-linsnalug-raw unchanged. The operation is CPU only, takes seconds, and is exactly reproducible from the five published or derived members.

Hardware is irrelevant to reproducing this checkpoint; the members were trained on 1x NVIDIA RTX 5090 (32 GB VRAM), 32 CPU cores, 92 GB RAM.

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).

For comparison on the same leaderboard: waxal-w2vbert-linsnalug-raw alone scored 0.742142 and the routed specialist pair scored 0.744835, so the average beats every one of its own ingredients. Only waxal-w2vbert-linsna-distilled, which continues from this checkpoint, scored higher as a single model, at 0.746787.

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 effectively seen Luganda: waxal-w2vbert-linsnalug-raw contributes Luganda exposure at weight 0.2 while the other four members were tuned on Lingala and Shona only, so Luganda behaviour should be assumed degraded relative to that checkpoint and is untested either way.

An honest caveat about the holdout figure above: it is Shona only, because that is the only holdout figure recorded for this checkpoint. Three of the five members are Shona-tuned or Shona-inclusive, so the average is expected to be stronger on Shona than on Lingala, and the single figure should not be read as an overall error.

This model is one component of the ensembles behind the submission. The 26-member character-level ROVER core scored 0.764915, well above this checkpoint's solo 0.746054, and the scored final submission, recipe p2n_mbr, selects per clip across seven such ensembles and scored 0.766791 public and 0.772552 private. This checkpoint appears in each core ensemble four times: once greedily decoded and three times re-decoded at blank penalties 1.0, 1.5 and 2.0.

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.