Abstract: ‍ ‍Decoding continuous neural intentions in the brain is a major challenge in brain-computer interface research. The physical quantity of speed， which has a natural continuum， is a feasible solution for decoding continuous neural intent， but there is still a gap in the current research on speed decoding in the field of brain-computer interface. In this paper， we propose a spontaneous speed imagery brain-computer interface paradigm and an accompanying multimodal neural signal decoding algorithm. This method uses a deep learning-based spatio-temporal feature attention network to decode continuous neural intentions， and achieves end-to-end decoding of multimodal data based on the extraction of local and global spatio-temporal features. In this paper， multimodal signals of left-hand clenched fist imagery were collected from 11 healthy subjects at 0 Hz， 0.5 Hz and 1 Hz， and the classification performance of the spatio-temporal feature attention network was verified using this dataset. The average classification accuracy and AUC values of the 11 subjects in the experiment were 89.6% and 99.0%， respectively. The experimental results show that the spontaneous speed imagery decoding using multimodal signals has the advantages of good performance and high efficiency， which is important for exploring continuous neural intention decoding in the brain and advancing practical applications of brain-computer interfaces.