Hello,
You need one noise covariance and one separate inverse model per subject. The noise covariance can be evaluated from the individual trials for each subject separately, as suggested here:
http://neuroimage.usc.edu/brainstorm/Tutorials/NoiseCovariance#Variations_on_how_to_estimate_sample_noise_covariance
If you are using the same anatomy, same selection of channels and same electrode positions for all the subjects, then the head model is the same for all the subjects. You can compute the model once and copy it to the other subjects. If you are looking for how to do this in a script, you can maybe find some inspiration in the script for the group analysis tutorial:
For the minimum norm solution (and its normalized versions sLORETA and dSPM), you need only a noise covariance matrix. The data covariance is needed only for the beamformer.
Francois