[Article by Nane Zietsman]
Dr Herman Kamper, senior lecturer in the Department Electrical and Electronic Engineering (E&E) at Stellenbosch University, is the recipient of the ISCA Award for the Best Research Paper Published in Computer Speech and Language (CSL) between 2016 and 2020. ISCA gives an annual award for the best research paper published in this journal over the last five years, and presents the award at the annual ISCA Interspeech conference. This year’s Interspeech took place in Brno, Czech Republic, and utilised a hybrid conferencing model to allow attendees from all over the world to join despite travel restrictions. Kamper attended and presented at the conference using their virtual platform.
The paper, titled: A segmental framework for fully unsupervised large-vocabulary speech recognition, draws on research that Kamper started during his PhD at the University of Edinburgh and continued with while doing his postdoctorate at TTI-Chicago. The paper is set against the backdrop of increased use in automatic speech recognition (ASR). “ASR is becoming much more prevalent in our everyday lives. Unfortunately, these systems are only available for a few well-resourced languages. This is because current ASR methods require thousands of hours of transcribed speech data,” Kamper explains. The paper proposes a new method to do speech recognition in the absence of transcriptions. According to Kamper, it is much easier to obtain speech recordings as opposed to transcriptions. “The proposed model learns word-like units directly from unlabelled speech audio. Since the model is not explicitly told what to recognise, it is an example of unsupervised learning, an important focus area within Machine Learning,” says Kamper.
Kamper’s PhD and postdoctorate research laid the foundation for work he has done with his students since joining The Faculty. “Although the idea and investigation presented in the CSL paper were good, the results indicated that there is still work to do. Despite significant developments, we still face challenges in building models that can identify recurring words in speech when they are not explicitly supervised to do so. We also struggle to build systems in languages where sources are limited, and there are still many questions surrounding unsupervised learning in general within Machine Learning.” According to Kamper, addressing these difficulties will provide more reliable speech technology and aid in better understanding continuous signals – a core focus within E&E. “Scientifically speaking, solving these problems can also tell us something about human learning. Humans learn many tasks, including language processing, without direct supervision. We are tasked with understanding how this works and how this process can be mimicked in machines,” explains Kamper.
The ISCA is one of the most influential organisations promoting research and teaching in the field of speech processing. Kamper’s paper will be amongst many outstanding writings published in the CSL journal. The CSL journal is one of the top three global journals investigating how Machine Learning can be applied to speech processing. Kamper describes the recognition of his work by the ISCA and CSL as amazing and humbling.
The Faculty of Engineering would like to congratulate Kamper on this achievement and look forward to his future research endeavours.
Photograph: Dr Herman Kamper (insert) received his ISCA Award via a virtual platform.