Speech Recognition Classification Leveraging a Spiking Neural Network Training Algorithm Based on Tempotron
Year: 2024
Type: article
Abstract: This paper examines the potential of Tempotron-based spiking neural network (SNN) training algorithms for speech recognition classification. Initially, the speech signal dataset was processed, with the signal decomposed into different frequency bands through Mel filter banks and the variations in speech intensity encoded using latency encoding techniques. These encoded features were then used as inputs, generating a feature map through the self-organizing map (SOM) unsupervised learning algorithm. This map reflected the similarity and topological structure between different features, providing a more abundant and abstract feature representation. Subsequently, for the task of speech signal classification, the paper established a two-layer SNN based on event-driven mechanisms, employing the improved LIF neuron model for information transmission. The network was then trained using the optimized Tempotron learning algorithm. Finally, the feature spike sequences extracted by the SOM were used as the input for the SNN. Five different weight update rules were employed in the Tempotron learning algorithm, and the resulting accuracy was recorded. The results of the experiment indicated that the combination of the Adam optimization algorithm with the Tempotron learning algorithm achieved a classification accuracy of 98.53%. This evidence demonstrates the efficacy of this paper in optimizing the Tempotron learning algorithm, which will inform future research aimed at enhancing classification results.
DOI: 10.54097/a4zpjb24
Authors: Erbin Ma, Wenxin Ge, Chongxuan Xiao
Institution: Northeastern University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4401666905