Dynamic imaging of the vocal tact is important for modeling speech through the acoustic-articulatory relation. The average duration of each sound is about 80ms. Movements of each articulator, in particular the tongue, should be captured with sufficient precision. Current clinical techniques use X-ray video fluoroscopy which involves ionizing radiation. Real-time MRI allows direct recording of speech motion [1] but is intrinsically limited in terms of resolution and SNR. Synchronization of MRI with an acoustic device is possible [2] but requires motion of vocal system to be highly reproducible. In this work we propose an optimized setup for achieving dynamic MRI of speech with high spatial and temporal resolution based on a combination of: an MR-compatible acoustic device allowing simultaneous recording of speech during MRI; and a retrospectively gated, motion-compensated image reconstruction that can deal with the variability of the subject repeating the same sentence over the acquisition.