Despite increasing global interconnectivity, radio remains a primary source of information for many communities in the developing world, including those in some parts of rural Africa. Because radio is accessible and can reach a large group of people, it is often the forum of choice in which a community will discuss topics of immediate concern such as healthcare, education, and service provision. Such public radio phone-in talk shows, hosted by small community radio stations, therefore contain important information that is not publicised in any other way, such as in the print media. By monitoring these radio conversations, information can be gathered that can be used to support humanitarian development and relief efforts by aid organisations. Valuable insights can be derived in this way such as indications on the spread of infectious diseases, the way people move during a disaster, how they perceive healthcare campaigns or their level of access to jobs and education.
In collaboration with the United Nations in Kampala, Uganda, researchers at the Electrical and Electronic Engineering Department (E&E Department) at Stellenbosch University developed an automatic speech recognition system that can automatically monitor the above-mentioned radio broadcasts and search them for indicative keywords or phrases. The overarching research question of the United Nations pilot study for which these systems were developed is: “Given the prominence of radio as a forum for public discussion, could this information be analysed to bring complete, timely and contextually rich insights that development practitioners can use?”1
The technology
The automatic speech recognisers that have been developed by the E&E Department are part of a system that continuously captures raw radio data. After automatic removal of most of the non-speech content, such as music, this is treated as a large and unstructured dataset.2 This dataset is processed by the speech recognisers to produce sets of possible text transcriptions which are searched for predetermined keywords or phrases. All potential hits are passed to human analysts, who listen to the audio and decide on its relevance.3 The result is a structured dataset of categorised text pertaining to the relevant topics. This process is detailed in the figure below.
Automatic speech recognition
Automatic speech recognition is at the front-line of this process. It converts the speech audio into a set of hypothesised transcripts. Unlike the raw speech, this text can be easily automatically searched. For the systems developed for the Ugandan languages, Luganda and Acholi, the speech recognition system can recognise vocabularies of 34,415 and 14,381 different words respectively.4 The team has found that, although the output of the speech recognisers contains many errors, important words are uttered often enough to be detected in this way. However, the uncertainty introduced by these transcription errors makes the final stage of the process, which involved human analysts, essential. This element of human involvement in a substantially automated process sets the system developed by the researchers apart from other similar approaches.
Human analysis
The final stage of analysis entails human analysts examining the original radio audio clips in order to remove false detections of topics and incorporate further semantic detail. This analysis has two stages: A rapid assessment, where original audio clips are extracted with matching filter rules, and labelling, where “the relevant fragments by categories (e.g. disaster.flood, health.disease-outbreak.cholera).5
Conclusion
The system developed by the E&E Department’s researchers is currently active and in continuous use by the United Nations in Uganda and will soon commence operation in Mali. It has already brought several important humanitarian issues to the attention of the aid organisation that would probably otherwise have been missed. The researchers are actively extending their work to new languages, with a special focus on languages that have so far never been technologically developed.
1 United Nations Global Pulse Using machine learning to analyse radio content in Uganda: Opportunities for sustainable development and humanitarian action (2017) 2.
2 United Nations Global Pulse Using machine learning to analyse radio content in Uganda: Opportunities for sustainable development and humanitarian action (2017) 6.
3 United Nations Global Pulse Using machine learning to analyse radio content in Uganda: Opportunities for sustainable development and humanitarian action (2017) 6.
4 United Nations Global Pulse Using machine learning to analyse radio content in Uganda: Opportunities for sustainable development and humanitarian action (2017) 9.
5 United Nations Global Pulse Using machine learning to analyse radio content in Uganda: Opportunities for sustainable development and humanitarian action (2017) 9.
