Cross-Lingual Transfer Learning for Enhancing Kinyarwanda Automatic Speech Recognition

In today’s voice-activated digital world, millions of people face a profound challenge: their native languages remain invisible to the very technologies that are transforming human-computer interaction. While speakers of major world languages like English, Mandarin, and Spanish enjoy seamless access to virtual assistants, transcription services, and voice-controlled systems across education, healthcare, and entertainment, countless indigenous and regional languages have been left behind in this technological revolution. Kinyarwanda, the vibrant national language spoken by over 13,246,394 million people across Rwanda, Burundi, Uganda, and the Democratic Republic of Congo, exemplifies this digital divide despite its crucial role in preserving cultural identity and facilitating daily communication, Kinyarwanda speakers are forced to abandon their mother tongue when interacting with modern speech recognition systems. This study uses cross-lingual transfer learning a new Speech-to-Text system formulated for Kinyarwanda, an under-developed Bantu language, with more than 14,104,965 million speakers mainly in Rwanda, Burundi, Uganda, and the Democratic Republic of Congo. The problem addressed involved merging different open-source Natural Language Processing data models with individualized preprocessing algorithms and acoustic feature extraction techniques. Using this innovative technique of combining Connectionist Temporal Classification with attention, assisted in achieving low Word Error Rate on standard Kinyarwanda speech corpora. The developed model is very efficient Automatic Speech Recognition system that can write spoken Kinyarwanda into text and promote digital inclusion and preserving linguistic heritage. This research demonstrates a developed complete speech recognition system through the deployment of the most recent deep learning architectures, such as Recurrent Neural Networks, Long Short-Term Memory models, and Transformer architecture. This new creation addresses the particular phonetic features, tonal differences, and morphological complexity of Kinyarwanda, while functioning within the confines of scarcity of training data as is characteristic of low resource languages.