10th U6+ Consortium of African Universities International Conference

Machine Translation (MT) systems typically rely on large parallel corpora to achieve high translation accuracy; however, Hausa remains a low-resource language due to the limited availability of digital linguistic and speech datasets. This study presents an optimized speech-to-speech Neural Machine Translation (NMT) framework for bilingual Hausa–English translation using a Bidirectional Long Short-Term Memory (BiLSTM) autoencoder integrated with the Differential Evolution (DE) optimization algorithm. Speech data were collected from literate Hausa speakers within the medical domain and pre-processed to improve the quality of the dataset. The BiLSTM architecture served as both encoder and decoder to capture contextual information from speech sequences, while a checkpoint mechanism enhanced training stability and convergence. The DE algorithm was applied to optimize the extracted features, reducing computational complexity and improving translation efficiency. Performance evaluation using BLEU score, Word Error Rate (WER), Character Error Rate (CER), and translation accuracy yielded scores of 0.45, 0.18, 0.16, and 83.2%, respectively. Comparative analysis with a baseline model demonstrated improved translation performance and computational efficiency. The results highlight the effectiveness of integrating deep learning and evolutionary optimization techniques to enhance speech-to-speech translation for low-resource language.