Now showing items 1-19 of 19

    • Activation gap generators in neural networks 

      Davel, Marelie H. (In Proc. South African Forum for Artificial Intelligence Research (FAIR2019), 2019-12)
      No framework exists that can explain and predict the generalisation ability of DNNs in general circumstances. In fact, this question has not been addressed for some of the least complicated of neural network architectures: ...
    • Code-switched English pronunciation modeling for Swahili spoken term detection 

      Kleynhans, Neil; Hartman, William; Van Niekerk, Daniel; Van Heerden, Charl; Schwartz, Rich; Tsakalidis, Stavros; Davel, Marelie H. (Procedia Computer Science: Spoken Language Technology for Under-resourced Languages, 2016-05)
      We investigate modeling strategies for English code-switched words as found in a Swahili spoken term detection system. Code switching, where speakers switch language in a conversation, occurs frequently in multilingual ...
    • Comparing grapheme-based and phoneme-based speech recognition for Afrikaans 

      Basson, Willem D.; Davel, Marelie H. (PRASA, 2012)
      This paper compares the recognition accuracy of a phoneme-based automatic speech recognition system with that of a grapheme-based system, using Afrikaans as case study. The first system is developed using a conventional ...
    • DNNs as layers of cooperating classifiers 

      Davel, Marelie H.; Theunissen, Marthinus W.; Pretorius, Arnold M.; Barnard, Etienne (arXiv:2001.06178v1 [cs.LG], 2020-01)
      A robust theoretical framework that can describe and predict the generalization ability of deep neural networks (DNNs) in general circumstances remains elusive. Classical attempts have produced complexity metrics that rely ...
    • Exploring minimal pronunciation modeling for low resource languages 

      Barnard, Etienne; Van Heerden, Charl; Hartmann, William; Karakos, Damianos; Schwartz, Richard; Tsakalidis, Stavros; Davel, Marelie H. (IOS Press Inc, 2015)
      Pronunciation lexicons can range from fully graphemic (modeling each word using the orthography directly) to fully phonemic (first mapping each word to a phoneme string). Between these two options lies a continuum of ...
    • Exploring neural network training dynamics through binary node activations 

      Haasbroek, Daniël G.; Davel, Marelie H. (Southern African Conference for Artificial Intelligence Research, 2020)
      Each node in a neural network is trained to activate for a specific region in the input domain. Any training samples that fall within this domain are therefore implicitly clustered together. Recent work has highlighted ...
    • Implications of Sepedi/English code switching for ASR systems 

      Modipa, Thipe I.; De Wet, Febe; Davel, Marelie H. (Pattern recognition association of South Africa (PRASA), 2013)
      Code switching (the process of switching from one language to another during a conversation) is a common phenomenon in multilingual environments. Where a minority and dominant language coincide, code switching from the ...
    • Improving the Lwazi ASR baseline 

      Van Heerden, Charl; Kleynhans, Neil; Davel, Marelie H. (Interspeech 2016, 2016)
      We investigate the impact of recent advances in speech recognition techniques for under-resourced languages. Specifically, we review earlier results published on the Lwazi ASR corpus of South African languages, and ...
    • Input parameter ranking for neural networks in a space weather regression problem 

      Lotz, Stefan; Beukes, Jacques P.; Davel, Marelie H. (In Proc. South African Forum for Artificial Intelligence Research (FAIR2019), 2019-12)
      Geomagnetic storms are multi-day events characterised by significant perturbations to the magnetic field of the Earth, driven by so-lar activity. Numerous efforts have been undertaken to utilise in-situ mea-surements of ...
    • Insights regarding overfitting on noise in deep learning 

      Theunissen, Marthinus W.; Davel, Marelie H.; Barnard, Etienne (In Proc. South African Forum for Artificial Intelligence Research (FAIR2019), 2019-12)
      The understanding of generalization in machine learning is in a state of flux. This is partly due to the relatively recent revelation that deep learning models are able to completely memorize training data and still perform ...
    • Medium-vocabulary speech recognition for under-resourced languages 

      Van Heerden, Charl J.; Barnard, Etienne; Davel, Marelie H. (SLTU, 2012)
      We report on the development of speech-recognition systems that are able to perform accurate recognition on mediumvocabulary tasks (i.e. tasks that require distinctions between approximately 200 different terms). We are ...
    • A neural network based method for input parameter selection 

      Lotz, Stefan; Beukes, Jacques P.; Davel, Marelie H. (Machine Learning in Heliophysics, 2019-09)
      NNs yield predictions, without aiding understanding of input-output relationship Fully connected networks mix signal from all inputs as information flows through the network Input parameter selection usually done
    • Predicting vowel substitution in code-switched speech 

      Modipa, Thipe I.; Davel, Marelie H. (Pattern Recognition Association of South Africa and Mechatronics International Conference, 2015)
      Abstract—The accuracy of automatic speech recognition (ASR) systems typically degrades when encountering codeswitched speech. Some of this degradation is due to the unexpected pronunciation effects introduced when ...
    • ReLU and sigmoidal activation functions 

      Pretorius, Arnold M.; Barnard, Etienne; Davel, Marelie H. (In Proc. South African Forum for Artificial Intelligence Research (FAIR2019), 2019-12)
      The generalization capabilities of deep neural networks are not well understood, and in particular, the influence of activation functions on generalization has received little theoretical attention. Phenomena such as ...
    • A smartphone-based ASR data collection tool for under-resourced languages 

      De Vries, Nic J.; Badenhorst, Jaco; Basson, Willem D.; De Wet, Febe; Barnard, Etienne; De Waal, Alta; Davel, Marelie H. (Elsevier, 2014)
      Acoustic data collection for automatic speech recognition (ASR) purposes is a particularly challenging task when working with under-resourced languages, many of which are found in the developing world. We provide a brief ...
    • Solar flare prediction with temporal convolutional networks 

      Krynauw, Dewald D.; Davel, Marelie H.; Lotz, Stefan (In Proc. South African Forum for Artificial Intelligence Research (FAIR2019), 2019-12)
      Sequences are typically modelled with recurrent architectures, but growing research is finding convolutional architectures to also work well for sequence modelling [1]. We explore the performance of Temporal Convolutional ...
    • The South African directory enquiries (SADE) name corpus 

      Thirion, Jan W.F.; Van Heerden, Charl; Giwa, Oluwapelumi; Davel, Marelie H. (Springer, 2019)
      We present the design and development of a South African directory enquiries corpus. It contains audio and orthographic transcriptions of a wide range of South African names produced by first-language speakers of four ...
    • Synthetic triphones from trajectory-based feature distributions 

      Badenhorst, Jaco; Davel, Marelie H. (Pattern Recognition Association of South Africa and Mechatronics International Conference, 2015)
      We experiment with a new method to create synthetic models of rare and unseen triphones in order to supplement limited automatic speech recognition (ASR) training data. A trajectory model is used to characterise seen ...
    • Text-based Language Identification of Multilingual Names 

      Giwa, Oluwapelumi; Davel, Marelie H. (IEEE, 2015)
      Text-based language identification (T-LID) of isolated words has been shown to be useful for various speech processing tasks, including pronunciation modelling and data categorisation. When the words to be categorised ...