Democratizing Machine Translation: A CPU-Centric Training Pipeline for Low-Resource Languages (A Kalabari Case Study)

Machine Translation (MT) systems for low-resource languages are scarce, particularly for highly divergent languages like Kalabari, a Niger-Congo language of the Ijo family. The persistent exclusion of these languages from modern language technologies is largely driven by a lack of parallel corpora, standardized tools, and the massive computational resources typically required for Neural Machine Translation (NMT). This research addresses this hardware bottleneck by detailing the end-to-end implementation of a foundational NMT system built entirely on consumer-grade CPU hardware. To establish this benchmark, a parallel corpus of 10,222 sentence pairs was manually created from available texts, carefully cleaned, and tokenized using SentencePiece Byte Pair Encoding (BPE) to mitigate morphological sparsity. A lightweight Transformer architecture was trained from scratch in OpenNMT-py on an AMD Ryzen 9 processor with 8GB RAM, utilizing SSD swap space and gradient accumulation to overcome memory limitations. The system achieved BLEU scores of 15.8 (Kalabari-to-English) and 13.5 (English-to-Kalabari), alongside promising chrF2 scores of 37.2 and 39.7 respectively. For deployment, the model was served on CPU using INT8 quantization via CTranslate2, a standard technique for reducing model size and accelerating inference on commodity hardware. The results demonstrate a reproducible, CPU-centric pipeline, proving that the lack of specialized GPU infrastructure is not an insurmountable obstacle for digital language preservation and baseline NMT development.