The Multilingual Speech Technologies (MuST) research group is a research niche area associated with the North-West University’s (NWU’s) Faculty of Engineering. It focusses on machine learning and statistical pattern recognition, as well as their application in speech technologies.

Over the years, the group has been involved in various speech recognition and speech synthesis development efforts for different languages, to transform a text into speech, and to recognise speech. MuST furthermore demonstrated the power of these technologies through integration into end-user applications, such as an automatic transcription system and a directory-enquiries system able to deal with the unique nature of South African accents and names.

Thanks to its extensive international networks, MuST researchers work with world-class experts on exciting projects.

In a project funded by the US intelligence agency IARPA, MuST was part of a consortium whose members included MIT and Johns Hopkins University. This international collaborative project focused on spoken term detection (recognising particular words from speech). MuST also worked with Google to create voices for four South African languages. The unit is also a node of the Centre for Artificial Intelligence Research, a South African research network that conducts research into various aspects of artificial intelligence.

Its speech-centred activities notwithstanding, MuST has a long-standing theoretical interest in the topic of generalisation: how information learned on a training set of samples is transferred to new inputs. The group is particularly interested in how generalisation functions in deep neural networks, the latter relating to a branch in computer science that uses statistical methods to help computers “learn” from data. Think image recognition: computer applications able to recognise objects in photos have acquired this capability from studying large sets of photos.

Internationally, deep neural networks have brought renewed energy and focus to the field of artificial intelligence, through a series of remarkable breakthroughs in fields as diverse as speech recognition, board games and self-driving cars. In each of these and many other applications, deep neural networks have reached previously unknown levels of accuracy, making human-level performance a distinct possibility. Interestingly though, many questions remain with regard to how deep neural networks achieve such good results, and how best to apply them.

In 2018, MuST has taken a better understanding of generalisation in deep networks as a new research challenge. Director Marelie Davel and co-research professor Etienne Barnard head the research team in pursuit of a deeper scientific comprehension of this important aspect of machine learning.

Given that deep neural networks have rapidly become a very important tool in current machine learning, an improved understanding of their application can have impact across multiple domains. “First and foremost we are intrigued and driven by these theoretical questions, but we also look forward to seeing our new work used in practical applications, not only in natural language and speech processing, but in the broader field of machine learning,” says Marelie.

MuST established a virtual reading group on the topic of deep learning that meets weekly. It consists of MuST researchers, postdoctoral fellows, postgraduate students and researchers from SANSA.