From a Multilingual Streaming ASR Backbone to Kenyan-Language Systems: Data-Centric Adaptation of Nemotron 3.5 for Kikuyu, Dholuo, and Kalenjin

This study investigates the practical engineering requirements for deploying streaming Automatic Speech Recognition (ASR) systems for three Kenyan languages: Kikuyu, Dholuo, and Kalenjin. Rather than training from scratch, the author adapts the NVIDIA Nemotron 3.5 ASR Streaming 0.6B model. The process utilizes a 'bridge' strategy, starting from a checkpoint already fine-tuned on Kenyan Swahili, and retains the model's cache-aware FastConformer RNN-T architecture to ensure true streaming capabilities.

A central theme of this work is that data quality is as critical as model architecture. The author details a rigorous pipeline involving corpus auditing, Unicode normalization, and duration filtering. A significant portion of the research focuses on the 'dirty' nature of real-world datasets, where inconsistent orthography, annotation artifacts, and missing audio often hinder performance. By documenting the specific cleaning decisions—such as why certain data was deleted rather than repaired—the study provides a transparent account of how these choices directly impact the supervision signal during fine-tuning.

The study successfully produced champion models for Kikuyu and Dholuo, which are now deployed in live streaming environments. Kikuyu achieved a 42.97% Word Error Rate (WER), while Dholuo reached 33.98% WER. The Kalenjin model is presented as an ongoing case study; its higher error rate (68.74% WER) and the documented failures during its development serve as a valuable record of the challenges inherent in low-resource language adaptation. The author emphasizes that aggregate WER is often insufficient for evaluating streaming systems, as it can mask issues like future-context leakage or unstable partial hypotheses.

This paper moves beyond simple leaderboard reporting to provide an auditable, end-to-end account of the ASR development lifecycle. By focusing on the 'how' of adaptation—including infrastructure, deployment, and error analysis—it offers a practical roadmap for researchers and engineers working with low-resource languages where public benchmarks may not exist or may not reflect real-world deployment constraints.