dronny111/lingala-shona-asr

Low-resource Lingala and Shona ASR training, decoding, and evaluation toolkit # Lingala and Shona ASR Low-resource speech recognition for Lingala and Shona, with an additional non-commercial Luganda route. The challenge metric is `1 - (WER + CER) / 2`, evaluated on 892 utterances without speaker or language metadata. ## System ```text language routing +-- Shona: w2v-BERT challenger with cross-model arbitration +-- Lingala: w2v-BERT epoch-2 low-LR continuation + transcript KenLM `-- Luganda: MMS-1B adapter + KenLM ``` Route sizes were 445 Lingala, 446 Shona, and one Luganda utterance. ## Repository structure | Path | Description | |---|---| | `pipeline/` | Training, decoding, evaluation, and submission entry points | | `src/` | Reusable model, CTC, language-model, metric, and training components | | `configs/` | MMS adapter configurations for Lingala, Shona, and Luganda | | `data/lm_nowiki/` | Transcript-only KenLM corpora and 5-gram binaries | | `tests/` | Unit, integrity, and release-boundary tests | ## Setup Python 3.11 is recommended. ```bash python -m venv .venv . .venv/bin/activate python -m pip install --upgrade pip python -m pip install -r requirements.txt ``` See `REPRODUCE.md` for data preparation, training, evaluation, and route composition. ## Development ```bash python -m pip install -r requirements-dev.txt ruff check . pytest ``` CI runs lint, builds the wheel, and executes the test suite on Python 3.11. ## License Original code is MIT licensed. Model, dataset, and language-model outputs retain their upstream terms. See `NOTICE.md` before redistributing derived models or outputs.