ravitejac/waxal-asr-solution

WAXAL ASR solution — retrained weights (Zindi code review)

Artifacts supporting the code review of submission GBc3jxZk in the Google WAXAL ASR Challenge (Zindi): public 0.757397 / private 0.766450, metric 1 − (WER + CER)/2, Lingala + Shona speech recognition. These checkpoints were retrained from scratch from the recipes in the reviewed code package; per-stage reproduction evidence ships with that package.

Contents

Path	Artifact	Base model	License of base
checkpoints/ft-7b-tta-spec/	7B decoder fine-tune (TTA specialist)	facebook/omniASR-LLM-7B (omnilingual-asr 0.2.0)	Apache-2.0
checkpoints/mwer-ttaspec-v2lr5/	final 7B (MWER expected-risk fine-tune)	facebook/omniASR-LLM-7B	Apache-2.0
checkpoints/w2vbert-lin-v1/, checkpoints/w2vbert-sna-v1/	per-language CTC fine-tunes	facebook/w2v-bert-2.0	MIT
corrector_v2/adapter/	LoRA r16 error-corrector adapter	google/gemma-4-E4B-it	Apache-2.0
lm/, lm_screen/	KenLM 5-gram binaries (participant-built text corpus)	KenLM (kpu/kenlm)	LGPL-2.1 (toolkit)
phase2_lid/	language-routing outputs (JSON)	produced with facebook/mms-lid-256	CC-BY-NC-4.0 (no model derivative included — outputs only)
submissions/	per-stage prediction files (participant model outputs)	—	—

Provenance and attribution

Fine-tuned checkpoints are derivatives of the base models above; credit to Meta AI (omnilingual ASR, w2v-BERT 2.0, MMS-LID) and Google DeepMind (Gemma). Base weights remain governed by their original licenses; no endorsement by the upstream authors is implied. Training data: the competition WaxalNLP corpus plus disclosed CC-BY-4.0 corpora (google/fleurs; AfriVoice mirrors KasuleTrevor/Lingala_100hrs, realtime-speech/shona1); full data-governance and external-data reviews are in the code package submitted to Zindi.

This repository contains no competition audio and no test-set reference transcriptions — prediction files are participant model outputs only.