The robustness to input variation of a spiking neural network model for speech recognition

No Thumbnail Available
Issue Date
Journal Title
Journal ISSN
Volume Title
The brain is capable of mapping auditory input to meaning. Regardless of the variability between the speech of different speakers or even the variability within the speech of a single speaker, the brain is able to understand it correctly. If this ability to recognise words and phones in varying conditions can be reproduced in a spiking neural network (SNN), it can be better understood how internal word and phone representations are learned by the brain. In this thesis I use an SNN with a biologically motivated unsupervised learning rule called spike timing-dependent plasticity (STDP), and homeostatic mechanisms that was shown to have a stable and balanced structure [1]. It was adapted to work with the temporal Spike-TIMIT dataset [2]. The main goal of this thesis was to test the robustness of this SNN against variations in the input data it receives. Two variations change the spatial structure of the spike pattern, by removing or switching part of the pattern. Another variation changes the temporal structure by jittering the spike times. I conclude that STDP learning does not clearly improve the robustness to recognise words and phones for this network. I also show that word recognition is more robust to variations in the input than phone recognition is.
Faculteit der Sociale Wetenschappen