Decoding speech imagery (SI) from electroencephalography (EEG) offers a promising avenue for Brain-Computer Interface but remains profoundly challenging due to the spectrally overlapping neural oscillations across multiple frequency bands ( \(\delta \) , \(\theta \) , \(\alpha \) , \(\beta \) , \(\gamma \) ), each encoding distinct linguistic processes. Most existing SI decoding methods lack effective multi-band fusion and computational efficiency. To address these issues, this paper proposes a Multi-Band Convolutional Tensor Network (MBCTN) that integrates parallel convolutional networks for band-specific feature extraction with a lightweight frequency band attention mechanism to dynamically weight the most discriminative features. Furthermore, by leveraging tensor train (TT) decomposition, MBCTN drastically reduces the number of model parameters while maintaining competitive performance. Comprehensive evaluations on the Chinese Imagined Speech Corpus (Chisco) demonstrate that MBCTN effectively captures cross-band interactions and achieves a decoding performance comparable to state-of-the-art methods, but with a fraction of the parameters.