Africa AI LedgerThe record of artificial intelligence across Africa

Africa AI Ledger · News

Speech AI Brings Women Farmers Into Advisory Systems

IFPRI and Farm Radio International are adapting speech recognition to real calls in African languages so women's questions can influence agricultural advice and programme design.

By Africa AI Ledger Desk · 4 November 2026 · Event: 4 November 2026 · 2 min read
Women farmers contributing voice messages through community radio to an African-language agricultural advisory network

Background generated in ChatGPT; Africa AI Ledger editorial template

Researchers working with Farm Radio International are using African-language speech recognition to make farmers' recorded questions more visible in agricultural research and advisory programmes.

The system, called Longa, converts voice messages into text that can be analysed at scale. IFPRI researchers have supported its development since 2021, adapting speech models with real calls from farmers rather than relying only on clean studio recordings. Their new account of the work focuses on whether women can participate meaningfully in the resulting services.

Farm Radio International works with more than 1,500 local radio stations and says its programmes reach over 200 million farmers in 38 sub-Saharan African countries and 68 local languages. Farmers can call toll-free numbers to ask questions, share experience and respond to broadcasts. The organisation holds about 8,000 hours of recorded speech that has been difficult to transcribe manually.

Voice matters because a smartphone application is not equally accessible to every farmer. Literacy, device ownership, data cost, travel and social expectations can all affect who participates. Radio and simple mobile calls build on channels that many communities already know and trust.

IFPRI reports that adapting models with field data improved performance by roughly 30 per cent across Luganda, Bambara and Swahili datasets. Fine-tuned models are available in Luganda and Bambara, with work expanding to other languages. The improvement is meaningful, but an average error reduction does not show whether women's speech, particular dialects or noisy environments receive equal accuracy.

Misrecognition can distort programme priorities. If the system repeatedly fails on certain speakers, their concerns may disappear from analysis even though they called. Gender-responsive evaluation should therefore compare error rates, topic capture and response quality across women and men, age groups, locations and recording conditions.

Consent and privacy also require more than removing a caller's name. Voice can reveal identity, location, health, finances and family circumstances. Farmers should understand whether recordings are used only to answer a question, to improve future programmes or to train models that may be shared with other organisations.

The strongest use of Longa is not automated advice without supervision. It is helping researchers and broadcasters listen to thousands of messages that would otherwise remain unexamined, identify recurring needs and adjust radio content or extension programmes. Human agronomists and local institutions should remain responsible for the advice delivered.

The project demonstrates a valuable principle for African AI: local data improve performance when they are collected through trusted institutions and governed around the people who produced them. It also shows why inclusion must be measured. Providing a call-in number is only the first step; women's questions must be accurately captured, answered and reflected in programme decisions.

Longa is a long-running research and implementation effort rather than a newly launched product. The latest evidence makes its contribution clearer: speech AI can widen participation, but only when language quality, privacy and institutional accountability are designed into the system from the start.

Source: International Food Policy Research Institute.

Source & attribution

Source: ifpri.org

Related coverage