A preliminary study of ConvMixer in the audio domain using Mel-spectrogram representation for lightweight speech command recognition

Audio classification is a core task for intelligent real-world systems, including voice control and environmental monitoring. Conventional Convolutional Neural Networks (CNNs) often involve substantial parameter redundancy, while Transformer-based architectures typically require higher computational overhead. In this study, we present an empirical investigation of ConvMixer, a simple yet powerful architecture originally developed for computer vision, for audio classification using Mel-spectrogram representations. To our knowledge, this work is among the earliest systematic efforts to adapt ConvMixer to the audio domain, bridging vision-inspired patch design and efficient speech understanding. The proposed framework incorporates a robust signal processing pipeline (denoising, WebRTC VAD, normalization) to generate fixed-size (128 \(× \) 32) Mel-spectrogram inputs. The ConvMixer model was configured with 8 deep blocks and 256 channels, with 719,117 trainable parameters (approximately 0.72M). On a Vietnamese speech command dataset of 4,460 samples and 13 classes, the ConvMixer-256/8 configuration achieved 96.75 ± 0.11 test accuracy and 0.9675 ± 0.0011 macro-F1 under multi-seed evaluation. To ensure fairness and statistical robustness in cross-model evaluation, we define a controlled benchmark protocol in which ConvMixer-256/8, ResNet-18, MobileNetV2, EfficientNet, and AST (tiny configuration) share the same preprocessing pipeline, Mel feature extraction settings, split policy, augmentation strategy, and epoch budget, with results reported as mean±std. Under this protocol, model comparison is reported using Accuracy, Macro-F1, loss, parameter count, inference latency, and memory footprint, with ConvMixer maintaining a compact 0.72M-parameter profile. Overall, these results indicate that a purely convolutional patch-based architecture remains a parameter-efficient and practical alternative for real-time speech command recognition, with clear potential for embedded and edge-AI deployment.