Abstract: This paper introduces ConvMixer, a highly efficient neural architecture designed for wake-word recognition on resource-constrained edge devices. By replacing heavy attention mechanisms with Multi-Layer Perceptron (MLP) "mixer" units and employing a five-stage curriculum learning strategy, ConvMixer achieves state-of-the-art performance in noisy, far-field environments. With only 119K parameters, the model rivals the accuracy of architectures nearly 50 times its size, providing a robust solution for small-footprint keyword spotting (KWS).
Introduction
Keyword spotting (KWS) is essential for hands-free device activation, yet it faces two primary hurdles: severe hardware constraints on edge devices and environmental noise in far-field scenarios. While large Transformer-based models handle noise well, their computational requirements are prohibitive for deployment on smart gadgets. Conversely, existing lightweight CNNs often fail when the Signal-to-Noise Ratio (SNR) is low. ConvMixer bridges this gap by combining a tiny parameter footprint with high noise robustness.
The ConvMixer Architecture
The core innovation of the model is the ConvMixer block, which facilitates multi-dimensional feature extraction without the overhead of self-attention.
As shown in Figure 1, the architecture utilizes:
- Multi-Dimensional Feature Extraction: 2D Depthwise Separable (DWS) convolutions capture frequency-domain patterns, while 1D-DWS convolutions handle temporal patterns.
- The MLP Mixer: Inspired by computer vision architectures, this unit uses simple MLPs to allow information to flow across temporal and frequency channels. This provides a "strong alternative to attention," achieving similar global feature interaction benefits with a significantly lighter memory footprint.
Curriculum-Based Multi-Condition Training
To enhance robustness, the authors employ a curriculum learning strategy. Rather than training on noisy data immediately, the model progresses through five stages of increasing difficulty:
- Clean audio only.
- Clean + 0dB SNR noise.
- Addition of -5dB SNR noise.
- Addition of -10dB SNR noise.
- Far-field simulation using Room Impulse Response (RIR) data to mimic echoes and distance.
This "tutor-like" approach allows the small model to master basic acoustic features before tackling complex environmental distortions.
Experimental Results
ConvMixer was evaluated on the Google Speech Commands V2 (12-label task) using MUSAN and BUT Speech@FIT databases for noise and reverberation simulation.
As detailed in Figure 3, the model demonstrates exceptional efficiency and robustness:
- Efficiency: With only 119K parameters and 22.2M MACs, it is significantly smaller than competitors like KWT-3 (5.3M parameters).
- Accuracy: It achieved 98.2% accuracy on the official V2-12 test set.
- Noise Robustness: In the most challenging -10dB SNR far-field test, the curriculum-trained ConvMixer (denoted as ConvMixer †) achieved 71.88% accuracy, outperforming MatchboxNet (62.35%) and performing comparably to the much larger KWT-3 model.
Ablation studies further confirmed that removing the MLP Mixer caused a ~7% drop in accuracy, highlighting its vital role in cross-channel feature interaction.
Conclusion
ConvMixer demonstrates that massive parameter counts and complex attention mechanisms are not strictly necessary for robust KWS. By leveraging efficient MLP-based mixers and a structured curriculum learning path, the model achieves high performance in noisy far-field conditions while remaining suitable for edge deployment. Future work will explore hardware-specific optimizations and scaling the architecture to larger vocabularies.
Summary updated