| Sumario: | This project builds on previous work where an earlier study sought to establish proof of concept on how speech recognition, along with natural language processing and text analysis techniques, can help to better identify farmer demands and target digital extension, particularly in the Global South.
The study ultimately led to development of Longa–an automated speech recognition tool for bantu languages and the work detailed in this report aimed to extend Longa’s capabilities by improving performance of previously tested Luganda models, extending the tool’s functionality to the West African context, with a focus on the Malian Language of Bambara, as well as creating a more user-friendly interface to allow for integration and deployment in more practical settings.
This report details the development of automatic speech recognition models for Luganda and Bambara designed for integration and operation within the existing framework at Farm Radio International (FRI), i.e. the Uliza application. The report is divided into four main sections describing activities and methods utilized in processing (annotation, preparation, cleaning, and modeling) radio recordings used to train the models, finetuning and evaluation of the models, as well as setting up interfaces on Hugging Face and Google Colab for testing.
|
|---|