A lightweight hybrid 2D–1D depthwise-separable CNN with frequency-only pooling for efficient small-footprint keyword spotting

Small-footprint keyword spotting (KWS) requires an effective balance between local time–frequency modeling and efficient temporal processing under tight deployment constraints. Rather than introducing a new primitive module, this work investigates a lightweight hybrid design that combines a shallow 2D time–frequency front-end with frequency-only pooling, a 1D depthwise-separable temporal backbone, and statistics pooling for segment-level aggregation. The design is motivated by a compute-conscious allocation of modeling roles: early local spectro-temporal feature extraction is handled by a shallow 2D block, while low-cost temporal context modeling is performed by 1D depthwise-separable convolutions. Using log-Mel spectrograms as input, the proposed model preserves temporal resolution through frequency-only pooling and summarizes frame-level representations with statistics pooling. On the Google Speech Commands v2 12-class task, the model achieves 97.5% top-1 accuracy with 54 K parameters and 3.8 M MACs. To characterize this operating point more rigorously, we report ablation and sensitivity analyses, reproduced baseline comparisons under a shared preprocessing and training protocol, on-device inference latency on Raspberry Pi 5 and LattePanda V1, and controlled white-noise evaluation across multiple SNR levels. The results show that the proposed configuration provides a competitive low-MAC operating point within the tested GSC v2 setting. In particular, the combination of a shallow 2D front-end, frequency-only pooling, and statistics pooling yields modest but consistent gains under repeated runs, while the lower computational cost also translates into favorable CPU inference latency. Under the tested white-noise protocol, the proposed model additionally shows smaller relative degradation than the reproduced baselines. These findings suggest that the studied hybrid design is a practical option for efficient on-device KWS when computational cost is a primary constraint.