Open Kalenjin Automatic Speech Recognition: Adapting Parakeet-TDT to a Low-Resource Nilotic Language

Abstract

Kalenjin, the Highland Nilotic language cluster of Kenya's third-largest ethnic community (6.36 million people in the 2019 national census [KNBS 2019]), has no openly available automatic speech recognition (ASR) model. We release the first: honest fine-tunes of NVIDIA's parakeet-tdt-0.6b-v3 (a 0.6B-parameter FastConformer encoder with a token-and-duration transducer decoder [Rekesh et al. 2023; Xu et al. 2023]), adapted to Kalenjin through a rebuilt SentencePiece tokenizer and reinitialized decoder and joint heads. We do not train from scratch, and we do not claim to beat the incumbent overall. Our second contribution is methodological: Kalenjin has no standardized orthography, and we show that this inflates Word Error Rate by seven to ten absolute points. Applying a linguistically grounded normalization (collapsing the voiced/voiceless stop contrast, which is non-phonemic in Kalenjin [Creider and Creider 2001]) moves our best model from WER 0.756 to 0.680 and the incumbent from 0.655 to 0.553, with no change to either system. The normalizer is a pure rule function requiring no lexicon, and we release it. Evaluation uses 198 leakage-controlled held-out unscripted utterances (156 Kipsigis, 42 Nandi) from AfriVoices-KE [Wanzare et al. 2026], scoring Character Error Rate as the primary metric with 95% paired-bootstrap confidence intervals (B=2000, seed=1234) [Bisani and Ney 2004]. Under matched greedy decoding our best fine-tune (v3 final) substantially leads the incumbent Whisper-Kln: greedy CER 0.2604 versus ≈0.483, both at the raw tier (L0). Overall the incumbent still leads, and we say so: with beam-5 decoding at the strongest normalization tier (L2), its CER of 0.1800 [0.1676, 0.1926] beats our 0.2310 [0.2223, 0.2404]. Training on the full 82,378-clip corpus only modestly edges a 30k-clip subset, with overlapping confidence intervals: corpus size was not the main lever. Our models are dialect-balanced (normalized CER: Kipsigis 0.2306, Nandi 0.2355). All 198 clips currently await native by-ear validation, so scores are directly comparable across systems but not yet formally citable.