This paper presents LISINet, a lightweight yet effective model for word-level sign language recognition. The proposed architecture comprises three main components: a convolutional block for spatial feature extraction, an LSTM-based module for temporal feature synthesis, and a fully connected layer for final prediction. Extensive experiments on three public benchmarks—LSA64, WLASL-100, and AUTSL—demonstrate that LISINet achieves competitive accuracy while maintaining a significantly low number of parameters and FLOPs. Although it does not rely on additional supervision or complex fusion strategies, LISINet outperforms several baseline and lightweight models in both accuracy and efficiency. These results highlight the potential of LISINet for real-time and resource-constrained applications in sign language recognition.