A Performance Comparison of 1D, 2D, and 3D CNN Architectures for Robot Voice Command Classification

Abstrak

This study presents a comparative analysis of one-(1D), two-(2D), and three-Dimensional (3D) Convolutional Neural Network (CNN) architectures for robotic voice command recognition using the Google Speech Commands dataset. Each architecture was evaluated in terms of classification accuracy, test loss, and computational efficiency to assess the trade-off between performance and resource demands. The experimental results show that the 3D CNN achieved the highest accuracy of 89.61% and the lowest test loss of 0.406, demonstrating superior capability in modeling spatiotemporal correlations within stacked spectrogram frames. The 2D CNN achieved an accuracy of 87.61% with balanced generalization and inference time. In comparison, the 1D CNN exhibited the lowest accuracy (68.90%) but the fastest inference speed (0.63 ms/sample), making it suitable for real-time robotic systems with limited computational resources. Qualitative evaluation confirmed that higher-dimensional CNNs yielded fewer misclassifications, especially for acoustically similar commands. Overall, the results indicate that the 2D CNN architecture provides the optimal compromise between accuracy and efficiency, while the 3D CNN offers the highest recognition capability. Future work will focus on developing lightweight 3D CNN or transformer-based models to enhance real-time performance in embedded robotic platforms.

Identifier

• OpenAlex: https://openalex.org/W7128520268
• DOI: https://doi.org/10.48084/etasr.14068