Comparative Analysis of CNN Architectures for Thai Accent Classification: VFNet vs. Sequential CNN

Abstract

This research presents a comparative study of Convolutional Neural Network (CNN) architectures for Thai accent classification. It contrasts a parallel architecture based on [1] VFNet: A Convolutional Architecture for Accent Classification, which uses multi-size filters simultaneously, with a sequential architecture that stacks different kernel sizes across layers (e.g., 3→5→7). The input features are Mel-Frequency Cepstral Coefficients (MFCCs) extracted from the Thai Dialect Corpus [2]. Experimental results show that both models achieve comparable accuracy and F1-scores. However, further analysis reveals that sequential models such as 5→5→5 and 7→5→3 outperform the VFNet-based parallel architecture in terms of lower parameter count and cross-entropy loss. A detailed 2D receptive field (RF) analysis also indicates that architectures with moderate RF sizes tend to deliver better classification performance compared to those with very small or excessively large RFs. These findings emphasize the practical advantages of well-structured sequential CNNs for real-world deployment under computational and memory constraints.