Radio is ubiquitous in the Global South, and it is a vital source of timely agronomic information for farmers in rural Africa, where one extension worker covers more than 4,000 farmers. Farm Radio International (FRI) supports over 700 radio shows in 40 countries across sub-Saharan Africa to broadcast discussion between experts, including CGIAR scientists, on locally relevant issues. FRI also receives voicemails so that it can provide locally relevant information. To overcome the challenge of listening and analyzing the overwhelming volume of voicemails, FRI has partnered with CGIAR to develop an AI-based solution to automatically extract insights from over 12 million listeners. This is contributing to reducing the digital divide in rural Africa.
Agronomic extension can be much more effective when it makes use of farmers’ knowledge and takes account of their needs and perspectives. Farmers themselves have a lot to say, and listening to them could enable much more effective extension. To this end, FRI has developed the Uliza system, which allows over 12 million listeners to call a number advertised during a show, which is then returned free of charge, allowing them to answer multiple-choice questions and leave a voice response to an open-ended question. The result is a huge repository of farmers’ knowledge and perspectives, disaggregated by age, gender, and country. Through one radio program, “On Air Dialogues,” for example, FRI received more than 12,000 responses, and 30% of calls were made by women. In their own words, farmers answered open questions such as “If you had the power to change things, what would you do to make a better life for farming families?” The project demonstrated that farmers have a lot to say, but analyzing and transcribing the large volume of calls has proved to be a challenge. About 90% of farmers’ responses were not analyzed and utilized in a timely manner.
“CGIAR’s innovative AI-based solution will empower Farm Radio to include millions of farmers in the conversation. By hearing each farmer’s voice, we can gain a deeper understanding of their perspectives and priorities. This valuable insight will lay the foundation for future projects with Farm Radio, ultimately enhancing our impact.” — Karen Hampson, Senior Manager (Program Development), Farm Radio International
To address this challenge, FRI has teamed up with the CGIAR Digital Innovation Initiative to create an automated solution that can rapidly analyze the content of voice messages and respond to farmers’ queries on air without the need for FRI staff to listen to each message individually. In 2022, using a human-centered design approach, the Initiative team identified FRI’s requirements and developed an initial prototype solution using a sample of audio data.
The new solution, published as a Technical Report, combines machine learning approaches in transfer learning, unsupervised learning, and corpus linguistics. The audio analytics approach typically relies on machine learning models trained using large databases of speech that have been manually analyzed and labeled. These databases do not exist for the majority of the 2,000 languages spoken in Africa, and the models would also struggle with local dialects, vocabulary, and speech patterns. The transfer learning method has addressed this data issue by first training a machine learning model based on a well-known language and then fine-tuning for local languages with a small amount of labeled data in the target language, resulting in an effective speech recognition tool. When applied to the FRI audio archive, the new prototype model, based on the fine-tuned XLR-S speech recognition model, outperformed other open-source speech recognition tools in both Swahili and Hausa.
Moving forward, the team will work on improving the transcription accuracy of the local languages by further fine-tuning the model with the full FRI audio archive. In 2023, the team aims to apply the solution to FRI’s new project, On-Air for Gender-Inclusive Nature-based Solutions, which addresses climate change and biodiversity loss issues through targeted radio programs broadcast from 220 stations across 38 sub-Saharan African countries. The Initiative will provide on-site capacity-building exercises for radio show hosts to familiarize them with the new solution. In 2024, the team plans to conduct a field study focused on evaluating the combined impact of radio shows and the automated tools on listeners’ food security, gender equality, and livelihood outcomes.
- Jones-Garcia, Eliot. 2022. Speech recognition, machine translation, and corpus analysis for identifying farmer demands and targeting digital extension. CGIAR Technical Report International Maize and Wheat Improvement Center (CIMMYT).
- Jones-Garcia, E., G. Kruseman, and B. Brown. 2021. Emotion classification and sentiment analysis for sustainable agricultural development: exploring available tools for analyzing African farmer interviews. CIMMYT Integrated Development Program Discussion Paper.
Header image: A woman in Sikilo village with her radio. In Senegal, climate forecasts and farming advisories are now available to more than 7 million rural people, via community radio stations. Photo by V. Meadu/CGIAR.