Optimization of Temporal Convolutional Networks Using Harris Hawks Algorithm for Efficient Embedded Speech Recognition

Abstract

Recent advances in embedded speech recognition require models that balance accuracy, efficiency, and low resource consumption. This paper presents a novel approach using Temporal Convolutional Networks (TCN) optimized by Harris Hawks Optimization (HHO) to enhance speech recognition performance on Raspberry Pi devices. The HHO algorithm dynamically tunes key hyperparameters of the TCN models such as convolutional filter sizes, dilation rates, and learning rates—in order to achieve high accuracy while reducing computational overhead. Experiments on benchmark speech datasets demonstrate that the proposed HHO-optimized TCN achieves higher recognition accuracy and faster inference compared to baseline models (e.g., standard LSTM, GRU). The work illustrates the potential of combining bio-inspired optimization with efficient convolutional sequence models to enable real-time speech recognition on resource-constrained edge devices.