Advancing AI Speech Recognition for Key Kenyan Languages

Automatic Speech Recognition (ASR) technology faces significant hurdles when applied to the diverse linguistic landscape of Africa, primarily due to issues like inconsistent orthography, data scarcity, and varied evaluation standards. This research presents a comprehensive engineering study focused on adapting NVIDIA's Nemotron 3.5 ASR Streaming model to develop high-performance speech recognition systems for three specific Kenyan languages: Kikuyu, Dholuo, and Kalenjin. The initiative builds upon an existing Kenyan Swahili-adapted checkpoint, leveraging its core architecture and streaming capabilities.

The methodology employed a rigorous data-centric approach, encompassing detailed corpus auditing, Unicode normalization to standardize text, and careful data splitting to ensure robust model training and evaluation. The researchers performed full-parameter fine-tuning, retaining the model's cache-aware FastConformer RNN-T and prompt conditioning. This meticulous process aimed to overcome common challenges in low-resource language ASR, such as managing missing audio, addressing speaker imbalances, and ensuring accurate evaluation in true-streaming environments.

Results indicate promising progress for Kikuyu and Dholuo, with word error rates (WER) of 42.97% and 33.98% respectively on internal evaluation sets. Character error rates (CER) were also reported, demonstrating the model's ability to transcribe these languages. While Kalenjin remains a work in progress with a higher WER, the study transparently reports both positive and negative findings, including challenges with non-speech labels and short-utterance over-generation, emphasizing an auditable account of the adaptation process rather than a state-of-the-art claim on public benchmarks.

This work is highly significant for Africa, particularly Kenya, as it directly addresses the critical need for advanced AI tools tailored to local languages. By enhancing ASR capabilities for languages like Kikuyu, Dholuo, and Kalenjin, it paves the way for more inclusive digital services, improved accessibility, and greater participation of local communities in the AI-driven economy. Such localized AI solutions are crucial for bridging the digital divide and fostering indigenous language preservation and development in the age of artificial intelligence.