Puno Quechua ASR — w2v-BERT 2.0 (scripted + silver spontaneous)
The system submitted to SIMBig 2026 Shared Task 1.
facebook/w2v-bert-2.0 with a 36-symbol character CTC head, warm-started from a
scripted-only checkpoint and fine-tuned on 18,168 scripted clips plus 6,442
spontaneous clips carrying silver labels. Greedy CTC decoding, no language model.
Development results, manual CTC decode:
| domain | CER | WER | references |
|---|---|---|---|
| scripted (2,284 clips) | 3.07 | 19.44 | gold |
| spontaneous (244 clips) | 18.69 | 36.19 | silver |
| out-of-domain (98 clips) | 6.12 | 40.45 | gold |
The spontaneous row is scored against the same automatic references the model was trained on, so it measures agreement with those labels rather than recognition accuracy.
Decode with the explicit CTC collapse in scripts/submit_qxp.py, not
tok.decode(skip_special_tokens=True), which inflates error by about 8.5 WER.
Use --max-sec 30: eleven spontaneous test clips run past 20 s, and a shorter cap
truncates the audio without truncating the reference.
Code, split manifests and per-utterance predictions: https://github.com/rmaacario/qxp-simbig2026