there are many easy extant ways to do voice coding. many models are released without a “voice embedding” model but they are easy to recreate by passing the gradients through the soft prompt