rekody/kalenjin-asr

Open Kalenjin ASR: evaluation harness, orthographic normalizer, and Parakeet-TDT fine-tuning pipeline (paper companion) # Kalenjin ASR Open Kalenjin automatic speech recognition: the evaluation harness, orthographic normalizer, and training pipeline behind the paper **"Open Kalenjin Automatic Speech Recognition: Adapting Parakeet-TDT to a Low-Resource Nilotic Language"** (Tony Kipkemboi, 2026). Preprint: ## Layout - `kalebench/` — the evaluation harness: scoring code (CER/WER, paired bootstrap, significance), the L0/L1/L2 orthographic normalizer (`scoring/normalize_kln.py`), the tier rescorer (`scoring/rescore_tiers.py`), the 198-clip held-out evaluation manifest (`tasks/asr/eval_unscripted.jsonl`, gold text + filename pointers, speaker IDs pseudonymized), and the archived evaluation artifacts every number in the paper is read from (`paper/artifacts/`). License: CC-BY-SA-4.0 (see `kalebench/LICENSE`). - `asr-train/` — the Parakeet-TDT fine-tuning pipeline (Modal): tokenizer rebuild, decoder/joint reinitialization, streaming training, evaluation entrypoints. License: MIT (see `asr-train/LICENSE`). ## Audio No audio is redistributed here. The evaluation manifest ships gold text plus a filename pointer; the audio remains in the gated source corpus (`Anv-ke/Kalenjin` on the Hugging Face Hub, CC-BY-4.0) under its own access and consent terms. Any use must cite Wanzare et al. (2026), AfriVoices-KE. ## Model checkpoints The fine-tuned checkpoints (v2 and v3 final) are published through the Rekody organization on the Hugging Face Hub; links are maintained here as they go live. - Rekody/whisper-large-v3-turbo-kalenjin: merged Whisper large-v3-turbo fine-tune (Kipsigis and Nandi). CER 0.1959 raw, 0.1629 orthography-normalized, on the frozen 198-clip held-out set. ## Reproducing the paper's numbers See the paper's Appendix A: every reported number maps to a command over the files in this repository, pinned by the frozen protocol in `kalebench/paper/artifacts/asr_eval_report.json` (B=2000, seed=1234, hashed evaluation set). ## Citing If you use this work, cite the paper: > Kipkemb …