Objective: To test the extent to which initial walking speed influences dual-task performance after walking intervention, hypothesising that slow walking speed affects automatic gait control, limiting executive resource availability. Design: A secondary analysis of a trial of dual-task (DT) and single-task (ST) walking interventions comparing those with good (walking speed ⩾0.8 m s−1, n = 21) and limited (walking speed <0.79 m s−1, n = 24) capacity at baseline. Setting: Community. Subjects: Adults six-months post stroke with walking impairment. Interventions: Twenty sessions of 30 minutes treadmill walking over 10 weeks with (DT) or without (ST) cognitive distraction. Good and limited groups were formed regardless of intervention received. Main measures: A two-minute walk with (DT) and without (ST) a cognitive distraction assessed walking. fNIRS measured prefrontal cortex activation during treadmill walking with (DT) and without (ST) Stroop and planning tasks and an fMRI sub-study used ankle-dorsiflexion to simulate walking. Results: ST walking improved in both groups (∆baseline: Good = 8.9 ± 13.4 m, limited = 5.3±8.9 m, Group × time = P < 0.151) but only the good walkers improved DT walking (∆baseline: Good = 10.4 ± 13.9 m, limited = 1.3 ± 7.7 m, Group × time = P < 0.025). fNIRS indicated increased ispilesional prefrontal cortex activation during DT walking following intervention (P = 0.021). fMRI revealed greater DT cost activation for limited walkers, and increased resting state connectivity of contralesional M1 with cortical areas associated with conscious gait control at baseline. After the intervention, resting state connectivity between ipsilesional M1 and bilateral superior parietal lobe, involved in integrating sensory and motor signals, increased in the good walkers compared with limited walkers. Conclusion: In individual who walk slowly it may be difficult to improve dual-task walking ability. Registration: ISRCTN50586966