Unpacking the Success of Riemannian Tangent Space Decoding for Speech Imagery

Speech Imagery (SI) is a highly intuitive Brain-Computer Interface (BCI) paradigm; however, while its popularity continues to increase, its current feasibility remains a subject of debate. In this work, we ’unpack’ the performance of Riemannian Tangent Space (TS) projection features classified with Logistic Regression (LR) to determine whether the pipeline’s efficiency in decoding SI from electroencephalography (EEG) reflects speech-imagery-specific activity or signals that merely covary with the task. Methods: We evaluated this pipeline on two open-access SI datasets that have fundamentally shaped the research field. We introduced a temporal sanity check to distinguish genuine task-related signals from the leakage induced by block-design paradigms. We then examined the spatial distribution of the LR weights, together with the independent components projecting onto the most heavily weighted channels, to assess the physiological origin of the discriminative features, and tested the pipeline’s invariance to drastically reduced feature sets to rule out ’shortcut’ learning from isolated sources. Contribution: We identify two sources of performance inflation that lead to a systematic overestimation of SI feasibility. First, evaluation strategies that ignore trial structure inflate accuracies, and sustained above-chance performance outside the imagery window shows that this leakage can mimic task-related features. Second, in the highest-performing condition the contrasted words differ in duration and in the visual cue presented throughout the trial, so that both muscular and cortical signals separate the classes without reflecting inner speech; a single channel and frequency band suffice to decode above chance. These findings prompt a reevaluation of feasibility claims in SI research.