Longa: An automated speech recognition tool for Bantu languages

Abstract

Farm Radio International (FRI) and the CGIAR Research Initiative on Digital Innovation have col laborated on the development of an end-to-end, automatic speech recognition pipeline for the tran scription, translation, and analysis of Swahili and Luganda. This task is particularly challenging due to the number of languages used by FRI's clients and the limited training data available for speech recognition in African languages. The tool is named 'Longa', or 'Let's chat' in Swahili. Longa will be used to answer the surplus of phone calls currently being received from smallholder farmers asking questions about radio programs which FRI does not presently have the capacity to address. When fully implemented, Longa should allow FRI to design their broadcasts more intricately in line with the needs of farmers and better deliver insights to those most in need, such as female and youth farmers. Key results from the collaboration include a series of design principles iteratively and col laboratively developed to reflect the common values and goals of FRI and the CGIAR, a proof of concept for Longa, building on open-source models and open access corpora, to be shared with the developer community upon completion of the final tool, a 10% improvement upon the state-of-the art automatic speech recognition in Luganda radio-speech performance and accuracy, some im provement in performance with audio enhancement processes using real-world data, and proof that fine-tuning is an effective approach to expanding Longa to new languages. The next steps of the collaboration will focus on the analysis and interpretation of an aggregation of farmer phone calls and integration with the existing FRI workflow and software.