The Makerere Radio Speech Corpus: A Luganda Radio Corpus for Automatic Speech Recognition

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1945–1954 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1945 The Makerere Radio Speech Corpus: A Luganda Radio Corpus for Automatic Speech Recognition Jonathan Mukiibi, Andrew Katumba, Joyce Nakatumba-Nabende, Ali Hussein, Josh Meyer Makerere University, Ronin Institute, Coqui Uganda, Egypt, USA jonathan.mukiibi@students.mak.ac.ug,{andrew.katumba, joyce.nabende}@mak.ac.ug, alizawahry1@gmail.com, josh@coqui.ai Abstract Building a usable radio monitoring automatic speech recognition (ASR) system is a challenging task for under-resourced languages and yet this is paramount in societies where radio is the main medium of public communication and discussions. Initial efforts by the United Nations in Uganda have proved how understanding the perceptions of rural people who are excluded from social media is important in national planning. However, these efforts are being challenged by the absence of transcribed speech datasets. In this paper, The Makerere Artificial Intelligence research lab releases a Luganda radio speech corpus of 155 hours. To our knowledge, this is the first publicly available radio dataset in sub-Saharan Africa. The paper describes the development of the voice corpus and presents baseline Luganda ASR performance results using Coqui STT toolkit, an open source speech recognition toolkit. Keywords: speech corpus, Luganda radio, automatic speech recognition 1. Introduction In sub-Saharan Africa, low internet penetration makes radio the most preferred medium of social communica- tion. Radio provides an opportunity for people’s con- cerns, particularly in rural communities, to get heard through the various radio talk shows where they can call in. Uganda has over 309 licensed radio stations, which creates a unique platform where views that could potentially harness the development of better policies are discussed (BBC, 2019). Therefore, there is a need to retrieve such valuable perceptions for national devel- opment. Previous work in the area of radio browsing using Au- tomatic Speech Recognition (ASR) has been done by the United Nations (Menon et al., 2018a). They have also experimented with Keyword Spotting (KWS) sys- tems in Uganda, and Somalia (Menon et al., 2018b). KWS for radio monitoring was developed as an alter- native to ASR systems due to the lack of a large corpus of transcribed radio data. In this case, the conventional approach of using ASR to perform speech-to-text and then search through the lattices for the presence or ab- sence of these keywords is not possible. In the last decade, the increase in the availability of large open-source speech datasets has propelled the application of deep learning in speech recognition re- search. As a result, research using various state-of-art ASR systems (Hannun et al., 2014) (Li et al., 2021) has produced better results compared to the traditional ma- chine learning approaches. However, the data demands of deep learning are well documented. Research on neural speech research for under-resourced languages is affected by the absence of speech datasets. On the other hand, this also frustrates the efforts to develop and adopt speech technologies in sub-Saharan Africa. Our target language in this research is Luganda, which is a Bantu …

Content truncated to 500 words